Using Social Media Networks for
Measuring Consumer Confidence:
Problems, Issues and Prospects
By
Jane-Vivian Chinelo Ezinne Igboayaka
Thesis submitted to the Faculty of Graduate and Postdoctoral Studies in partial fulfillment of
the requirements for the degree of
Master of Science
in
Electronic Business Technologies
Faculty of Engineering
University of Ottawa
Ottawa, Ontario
© Jane-Vivian Chinelo Ezinne Igboayaka, Ottawa, Canada, 2015
ii
Abstract
This research examines the confluence of consumers’ use of social media to share information
with the ever-present need for innovative research that yields insight into consumers’ economic
decisions.
Social media networks have become ubiquitous in the new millennium. These networks,
including, among others: Facebook, Twitter, Blog, and Reddit, are brimming with
conversations on an expansive array of topics between people, private and public organizations,
governments and global institutions. Preliminary findings from initial research confirms the
existence of online conversations and posts related to matters of personal finance and
consumers’ economic outlook.
Meanwhile, the Consumer Confidence Index (CCI) continues to make headline news. The issue
of consumer confidence (or sentiment) in anticipating future economic activity generates
significant interest from major players in the news media industry, who scrutinize its every
detail and report its implications for key players in the economy. Though the CCI originated in
the United States in 1946, variants of the survey are now used to track and measure consumer
confidence in nations worldwide.
In light of the fact that the CCI is a quantified representation of consumer sentiments, it is
possible that the level of confidence consumers have in the economy could be deduced by
tracking the sentiments or opinions they express in social media posts. Systematic study of
these posts could then be transformed into insights that could improve the accuracy of an index
like the CCI. Herein lies the focus of the current research—to analyze the attributes of data
from social media posts, in order to assess their capacity to generate insights that are novel
and/or complementary to traditional CCI methods.
The link between data gained from social media and the survey-based CCI is perhaps not an
obvious one. But our research will use a data extraction tool called NetBase Insight Workbench
to mine data from the social media networks and then apply natural language processing to
analyze the social media content. Also, KH Coder software will be used to perform a set of
statistical analyses on samples of social media posts to examine the co-occurrence and
clustering of words. The findings will be used to expose the strengths and weaknesses of the
iii
data and to assess the validity and cohesion of the NetBase data extraction tool and its suitability
for future research.
In conclusion, our research findings support the analysis of opinions expressed in social media
posts as a complement to traditional survey-based CCI approaches. Our findings also identified
a key weakness with regards to the degree of ‘noisiness’ of the data. Although this could be
attributed to the ‘modeling’ error of the data mining tool, there is room for improvement in the
area of association—of discerning the context and intention of posts in online conversations.
iv
Acknowledgements
Above all, I want to thank God Almighty, for the strength, good health and patience He gave
me through the course of this thesis write up.
I want to express my sincere gratitude to Professor Kindra, my supervisor for his continued
support even though it has not been an easy journey.
I would like to thank my co-supervisor, Professor Mulvey who has motivated me all through
writing this thesis. His patience, support, understanding, persuasion, knowledge, guidance and
encouragement helped me push through even when I had given up. I want to say thank you.
I also want to acknowledge my husband, Ifeanyi who has given me enormous and utmost
support—moral, spiritual and physical, love, guide, encouragement, motivation and continuous
prayers in making sure I achieve my dreams and pursue my goals to reach the next phase of
my life.
My sincere thanks go to my parents, Engr. (Sir.) Martin and (Lady) Evelyn Igboayaka for
seeing me through my education and life. Their unrelenting prayers, love and support—in every
way that gave me the inspiration to finish this research and thesis. Also, I want to thank my
siblings; Sandra, Jennifer, Martin, Marcel and Sophia Igboayaka for their prayers and support.
Last but not the least, I want to thank all my friends and well-wishers; Chinenye, Edidiong,
Susan, Chidi, Joshua, Theresa, Chinedu, Chrestina, Akhere, Chidinma, Dafe, Abdul, Uzoma,
Oduwa, Ebere, Chinweike, Dorcas and Chigozie for their encouragement and support during
the period of this thesis write up.
v
Table of Contents Abstract .................................................................................................................................................. ii
Acknowledgements .............................................................................................................................. iv
Table of Contents .................................................................................................................................. v
Table of Figures................................................................................................................................... vii
List of Acronyms .................................................................................................................................. ix
Chapter 1 Introduction ......................................................................................................................... 1
1.1. An introduction to established methods of Consumer Confidence Index measurement ........ 4
1.2. Illustration of the use of Consumer Confidence Indices ......................................................... 4
1.3. Social Media Networks as a source of data for Consumer Confidence measurement ............ 6
1.4. Research Objective ................................................................................................................. 6
1.5. Research Question .................................................................................................................. 6
1.6. Research Motivation ............................................................................................................... 7
1.7. Research Approach and Contribution ..................................................................................... 7
1.8. The structure of thesis ............................................................................................................. 9
Chapter 2 Literature Review ............................................................................................................ 10
2.1. The economic indicator of consumer confidence index ............................................................ 10
2.3. Measurement of Consumer Confidence ................................................................................ 18
2.4. Social media networks .......................................................................................................... 19
2.5. Attributes of Social Media Networks .................................................................................... 22
2.6. Data mining concepts ............................................................................................................ 24
2.6.1. Natural Language Processing concept .......................................................................... 26
2.6.2. The Linguistic Inquiry and Word Count concept ......................................................... 28
2.6.3. Life History Calendar (LHC) ........................................................................................ 28
2.6.4. Method Proposed by Nigam and Hurst ......................................................................... 29
2.7. Data mining tool ................................................................................................................... 30
2.7.1. About NetBase® ............................................................................................................ 30
2.7.2. Advantages of Using NetBase® .................................................................................... 31
2.7.3. Limitations of NetBase® ............................................................................................... 31
2.8. Chapter Summary ................................................................................................................. 32
Chapter 3 Methodology ..................................................................................................................... 33
3.1. Social media data collection ................................................................................................. 33
3.2. Keywords for search ............................................................................................................. 34
3.3. Research data analysis method ............................................................................................. 36
3.3.1. KH Coder ...................................................................................................................... 37
3.3.2. Raw dataset classification criteria ................................................................................. 39
vi
Chapter 4 Results and Analysis ....................................................................................................... 41
4.1. Assessing the robustness of social media networks data ...................................................... 45
4.1.1. Part A: Qualitative analysis based on Infographics ...................................................... 45
4.1.2. Part B: Quantitative analysis of aggregated summaries ................................................ 60
Tax: Top terms ......................................................................................................................... 61
4.1.3 Part C: An analysis of SMNs content contribution ....................................................... 78
4.1.4 Part D: Analysis of raw data indexed by NetBase ........................................................ 83
4.2. KH Coder: Exploration of natural attributes of raw data indexed by NetBase ..................... 87
Chapter 5 Summary & Conclusions................................................................................................. 91
5.1. Summary .................................................................................................................................... 91
5.2. Conclusion ................................................................................................................................. 93
5.3. Future Study ............................................................................................................................... 95
Bibliography ........................................................................................................................................ 96
vii
Table of Figures Fig. 1: A comparative illustration of the correlation between the CCI and other economic indicators .................. 5 Fig. 2: Example of the architecture of a typical opinion mining application used for a product review (Tsytsarau
and Palpanas 2012) ........................................................................................................................................ 8 Fig. 3: Differences between the three oldest forms of consumer confidence index measurement (Source: Reserve
Bank of Australia) ........................................................................................................................................ 11 Fig. 4: A survey of restaurant customers on their assessment of their personal finances ..................................... 14 Fig. 5: A global CCI of sorts generated from a survey of 60 countries for the period of October – December
2013; Source: Nielsen, Consumer Confidence Concerns and Spending Intentions around the World
Quarter 4, 2013. ........................................................................................................................................... 19 Fig. 6: Global Penetration of social platform-Account ownership and Active usage— (Chaffey, Dave, 2014) .. 24 Fig. 7: Illustrating the association of words with other words used with them, an example of co-occurrence
network of words. Source: (Higuchi, 2013) ................................................................................................. 38 Fig. 8: Illustrating the association of words with correspondence analysis. Source: (Higuchi, 2013) .................. 38 Fig. 9: An example of output from hierarchical analysis using KH Coder. Source: (Higuchi, 2013) .................. 39 Fig. 10: Word cloud of (a) Top Emotions expressed; (b) Top Terms used and (c) Most used hash tags .... 41 Fig. 11: Qualitative summary attributes of data associated with Tax keyword .................................................... 48 Fig. 12: Qualitative summary attributes of data associated with the Capital Expenditure keyword ......... 52 Fig. 13: Qualitative summary attributes of data associated with the Employment Condition keyword ............... 55 Fig. 14: Qualitative summary attributes of data associated with the Health keyword .......................................... 56 Fig. 15: Qualitative summary attributes of data associated with the Consumer Confidence keyword ................. 57 Fig. 16: Qualitative summary attributes of data associated with the Political Affiliation keyword ...................... 58 Fig. 17: Qualitative summary attributes of data associated with the Retirement keyword ................................... 59 Fig. 18: Graphed Top term representation of the quantified attributes of data from social media networks,
associated with Tax ...................................................................................................................................... 63 Fig. 19: Quantified representation of the top emotions expressed that are related to Tax .................................... 67 Fig. 20: The hashtags used in conversations that are contextually associated with Tax ....................................... 69 Fig. 21: The volume of top terms associated with Capital Expenditure ............................................................... 73 Fig. 22: Graphical representation of the number of Top emotions expressed, relative to Capital Expenditure .... 76 Fig. 23: The most used hashtags related to the subject of Capital Expenditure .................................................... 77 Fig. 24: The top social media networks in terms of the volume of conversations ................................................ 82 Fig. 25: Co-occurrence associations of indexed raw datasets ............................................................................... 88 Fig. 26: Relationship between frequency of terms and the number of paragraphs in a set of data ....................... 89 Fig. 27: Hierarchical cluster of terms contained in the raw data set indexed by NetBase® .................................. 90
viii
List of Tables Table 1: Summary of Consumer Confidence Indexes from around the world………………………………..…15
Table 2: A frequency count summarizing the Top terms & Top emotion words related to Capital
Expenditure……………………………………………………………………………………………………....42
Table 3: The third output from a NetBase search; the indexed dataset extracted from social media networks for
the search term of Capital Expenditure…………………………………………………………………..………44
Table 4: Illustration of the how niche social media networks could serve some subject area better than the
others…………………………………………………………………………………………………………..…78
Table 5: Indexed raw data from NetBase associated with Tax for 12 month period………………………….…85
Table 6: Part B of Table 5 showing the sampling interval……………………………………………….………86
ix
List of Acronyms
CBC Conference Board of Canada
CC Consumer Confidence
CCI Consumer Confidence Index
CCS Consumer Confidence Survey
IR Information Retrieval
LIWC Linguistic Inquiry & Word Count
MS Excel Microsoft Excel
NLP Natural Language Processing
PAYE Pay-As-You-Earn
POS Part-of-Speech
SM Social Media
SPSS Statistical Package for the Social Sciences
WSD Word Sense Disambiguation
_____________________________ 1 Bloomberg News of Nov. 28, 2014 illustrates the associated importance of consumer confidence in the
article ‘U.K. Consumer Confidence Stalls as Economic Outlook Deteriorates’ (Ryan, J. - Bloomberg,
2014)
1
Chapter 1 Introduction
A notable indication of the economic prospects of an economy is the level of its consumers’
confidence in the economy. In mainstream economic discussions, this is expressed as a
quantified metric called the Consumer Confidence Index (CCI). It is a crucial data that it is
referenced by businesses, governments and other institutions when they make announcements
about the strategy direction their institution is taking or is about to pursue. In mainstream
business media channels such as Bloomberg1, time and again, the CCI is referenced and makes
headline news.
The confidence consumers have in their ability to spend on goods and services is a cue of the
direction of economic activities, whether they feel they have or will have the economic means
to buy goods and services they require. These in turn influence the manufacturing of goods and
delivery of services, affecting the levels of labor employment, the income tax generated for the
government, profits or loss enjoyed by businesses and the commitment they go on to make on
capital investment.
It could be argued that the historic reference to the CCI as an economic indicator has been time
tested and thus is a ‘reliable’ way of telling whether economic activity will grow or decline in
the near future. For example, the Conference Board of Canada has been publishing a CCI since
1980. During these past 35 years, despite the creation of other metrics to measure the
consumers’ confidence, the CCI is still referred to each time it is published.
Nonetheless, the reliability of the consumer confidence index, is directly linked to how
accurately its methods of measurement capture the economic sentiment of the population of a
country. In this regard, the methods of measurement start with a manual random survey of
different demographics of the population. These surveys are known as Consumer Confidence
Surveys (CCS) and they are issued by postal mail, electronic mail & telephone campaigns.
2
As retail technology has evolved, the measurement of consumer confidence has extended to
including data from retailers’ point-of-sales systems. This data is used to synthesize trends of
consumers’ buying pattern (Hudszinki, 2014), which is then used to infer the implicit
confidence of consumers, based on how much they are spending relative to historic patterns.
Complementarily, this information is combined with the conventional questionnaires, to give
an indication of the general economic prospects. The recently published CCI of the United
States was generated using probability design random sample (The Conference Board, 2014).
It shows that consumer confidence stands at 88.7% for November, 2014 down by 5.1%
compared to October,.
These variant but convergent methods of measurement tend to improve the accuracy of the CCI
they are used to generate. On one hand some of these methods look to elicit the consumers’
opinions on their economic situation—from the preceding six months to the six months ahead.
This is the classic CCS method. Also, technology driven methods are used to monitor
downstream spending patterns of consumers across vast retail horizontals. The diversification
of the methods of CCI measurement can only have a positive influence because they entail
looking at the subject of interest from more than one perspective.
When combined, these evolving means of measurement are destined to improve the accuracy
of the CCI, from the researcher’s point of view. Furthermore, following the advent of social
media networks, new opportunities to generate more insights into the sentimental bias of
consumers towards the economy continue to open up.
Using social media networks as a platform for social intelligence gathering is catching on as a
data source in the market intelligence research industry. Companies such as Nielsen, Netbase,
Frost & Sullivan, and Mintel enlist their use of content from social media networks, for
intelligence gathering and marketing purposes (Netbase, 2010) (Nielsen, 2013). It is no surprise
that a lot of research continues to concentrate on the improvement of analytical methods and
frameworks to translate such content into useful information.
It is this translation that is pivotal to the usefulness and correctness of whatever information is
synthesized from social media networks. For example, an area of study like sentiment analysis
frameworks has originated concepts such as Natural Language Processing (NLP) that can be
used as a tool to translate data from social media networks into meaningful insight.
3
In review, the indicator, CCI and how it is measured have expanded their scope beyond the use
of questionnaires alone, and data sources present opportunities to gain another perspective on
consumer behavior in terms of: economic habits (past), situation (present) or intentions
(future). At this time, social media networks are a relevant sources for the above.
However, not a lot of work has been done to assess the characteristics of the data extracted
from social media networks and the emerging trends reflected by and resulting from this data.
Therefore, it is the goal of this research to explore at a macro level any and all trends that can
be captured from the SMNs extracted data, using a commercially available social intelligence
gathering tool.
The purpose of undertaking this research study is to explore the attributes of data from SMNs
and assess how robust they are for the purpose of synthesizing a metric like CCI. This research
would also help novice and established professionals in the area of electronic business
technologies and data mining in particular to evaluate the issues and uncover the prospects of
utilizing such data in research studies.
Therefore the scope of this thesis pertains to the use of a social intelligence gathering tool to
extract data from social media networks. A qualitative analysis is performed on the data to
evaluate such attributes as ‘representative’ of popular opinion during a period of interest. To
do so, graphical word cloud summaries synthesized from the data are analyzed in order to see
whether they give a hint about the economic sentiment of consumers.
In this introductory chapter, conventional methods used to collect data for the synthesis of a
CCI are reviewed in Section 1.1. In Section 1.2, a brief illustration of how the CCI compares
to other economic indicators is presented. In Section 1.3, social media networks and their
attributes are briefly discussed, following the wave of research interest in them as a revealing
source of the aspects of human psychological perspectives in the digital age.
Section 1.4 gives a detailed account of the objectives of this research and Section 1.5 lays out
the methodology of research implementation. Finally, the structure of the remainder of the
thesis is presented in sections 1.6 through 1.8.
4
1.1. An introduction to established methods of Consumer Confidence Index
measurement
The conventional means of determining consumer confidence are different from one country
to the other. Also, the organizations that provide this information are different, being either a
not-for-profit think tank or government departments and sometimes profitable private
businesses.
In Canada and the United States of America, telephone interviews and postal mail
correspondences are used to survey a random selection of households. This could range from
2,000 to 5,000 correspondents (The Conference Board of Canada, 2014). The Conference
Board which publishes the CCI for the United States posits five questions on the CCS
questionnaire sent to corresponding households (McKinsey 2012). They are:
Current business conditions?
Business conditions for the next six months?
Current employment conditions?
Employment conditions for the next six months?
Total family income for the next six months. Survey participants are asked to answer
each question as "positive," "negative" or "neutral?”
The above questions are carried out at the following intervals: every two months and at the end
of each quarter (The Conference Board, 2011). In recent times however, the Conference Board
of Canada has partnered with Nielsen, a global market intelligence company in the retail space
to provide its Consumer Confidence Index. According to the information on the CBC’s
website, Nielsen uses a technique known as probability design random sample on the primary
data. Whether this primary data is solely based on conventional survey methods or data from
other sources like social media networks is not explicit.
Other methods used to measure consumer confidence are predominantly those used for the
survey methodology. Nonetheless, there is an increasing use of supplementary data from other
sources such as retail point of sales and social media networks. In Section 1.2, the discussion
is about how Consumer Confidence Index as an economic indicator is used.
1.2. Illustration of the use of Consumer Confidence Indices
CCI information influences business strategy decisions made by manufacturers, retailers, banks
and governments, relative to other key economic indicators (Investopedia, 2014). A
consistently declining trend in CCI would indicate that consumers have a negative economic
5
outlook which implies further decline in economic activity. While a positive CCI trend would
be taken as antecedent to growth in economic activity in the near future—good for business.
For example, Fig. 1 attempts to show the correlation between different economic indicators in
the United Kingdom and the level of the CCI over an unspecified period of time. This is one
example of how businesses could base their decision of either investing in production capacity
or cut labor costs, based on the anticipated economic activity inferred from the level of
consumer confidence.
Similarly, banks can also take measures to curb the anticipated increase in borrowing by
consumers and the use of credit cards. The government could come up with measures such as
reduction of taxes in order to encourage spending. The opposite will apply in a case of an
upward trend in the index (higher production/inventories and tax increase).
From 1921 to 1933, the United States government adopted a policy of laissez-faire which
required them to be less involved in the affairs of the economy (however, there was no CCI at
this time). They achieved this by keeping taxes low thereby encouraging businesses to invest
more money for expansion and also creating more money for consumers to spend (BBC, 2014).
Fig. 1: A comparative illustration of the correlation between the CCI and other economic indicators
Therefore it seems reasonable to conclude that CCI influences decisions made by pivotal
economic entities such as governments. One aspect of such dynamic has to do with the
emergence of new sources of data that could be used to measure consumer confidence.
6
1.3. Social Media Networks as a source of data for Consumer Confidence
measurement
In an era of social media networks facilitated by the advances of the internet, its components
including blogs, Facebook®, Twitter®, FourSquare® et al, are a versatile source of
information which could be harnessed for assessing the human perceptions that reflect their
confidence in the economy. For example, Facebook and Twitter are popular networks used to
express opinion, a state of mind or a perspective which could be mined for the sentiment of
users relative to their economic situation. FourSquare is used to identify with geospatial
preferences, such as where an individual is geographically located.
Also, blogs and the commentary they generate are a rich source of opinion bias towards a
subject of discussion; e.g. an election, a new movie, a new government policy, stock market
fluctuations, a commercial investment deal and much more. Aggregation of such information
can be used to ‘know’ the economic state of consumers.
Furthermore, 60% of the world’s population has access to the internet and now interacts using
social media networks (Barry et al. 2011). Therefore tapping into this ‘new’ data source could
give a supplementary boost to the conventional methods of data gathered to measure the level
of consumer confidence in the economy.
How the richness of such data could benefit how information such as consumer confidence
could be measured is one aspect this research aims to explore. In the next section the objective
of the research is articulated.
1.4. Research Objective
Our research will explore first how representative past social media analytics data is/are as a
source of information about the economic sentiment of consumers. Next, the study will analyze
the intrinsic attributes of the data to assess their strengths and weaknesses relative to the level
of consumer confidence in the economy.
1.5. Research Question
The research question is: is data from social media networks suitable and reliable for
measuring CCI? Our research seeks to explore the attributes of data from social media
networks to assess their robustness for use in the measurement of CCI.
7
1.6. Research Motivation
For the past ten years, electronic business technologies continued to transform the landscape
of different industries, professions, politics, religion and the global economy. It is known that
we are in a ‘data-driven’ economic era, where timely knowledge is vital.
This reality has prompted the emergence of new opportunities to apply e-business concepts in
non-traditional areas. The measurement of CCI from social media networks is one such area.
The research student believes that by applying e-business tools and concepts, this research
would facilitate an understanding of how to ‘mine’ data from unstructured sources like social
media networks, develop further an ability to analyze such data and critically evaluate the
usefulness of the data for application in a research area of interest.
The research will examine
i. How to use technological tools to exploit hidden insights in unstructured data from
social media networks, in addition to the underlying concepts upon which such tools
are built.
ii. The research would explore the versatility and suitability of social media networks data
for measuring CCI.
iii. How to undertake data analysis and design models for macro information origination
from data.
iv. Challenges of data gathering, data management and administration for research
purposes.
1.7. Research Approach and Contribution
There are already tools that mine datasets for indicative trends such as opinion or sentiment.
For example, General Inquirer from Harvard University, SAL—Sensitive Artificial Listener,
TS—Twitter Sentiment, RA—RateItAll are examples of opinion mining retrieval tools
(Tsytsarau & Palpanas, 2012). The limitation of these tools for purposes of this study are
coming from the models upon which they are based, because the implicit assumptions may be
unsuitable for the exploratory direction of this research. It is for this reason that a non-
specialized SMNs data gathering tool was of interest. After studying available tools, NetBase®
was identified as a suitable tool for this study, because its design is such that it can extract data
from all SMNs that can be accessed through ‘crawling’ algorithms on the internet. It is not only
8
able to extract the exact conversations (Soundbites) from these SMNs but it deduces summaries
about the data using various aspects of the NLP model it implements. It also makes the raw
Soundbites data accessible for further analysis. The fact that NetBase® does not focus on the
feed of a specific SMN and makes the extracted SMN data available for further analysis, is
why it was selected for this research. It is recommended for other researchers who want to work
with data from as many SMNs as possible, even though they will have no control over selection
of specific SMNs.
NetBase® needs keywords as a guide to extract SMNs data that are relevant to the subject of
interest. Therefore, the selection of the keywords was critical to minimize the extraction of
Soundbites that are not connected to CCI. How it functions is illustrated in Fig. 2. To decide
on the keywords to use, mainstream news media reactions and discussions were evaluated,
after each publication of CCI in North America and Europe. Keywords such as ‘capital
investment’, ‘retirement’, ‘pension’, and ‘capital expenditure’ identified on this basis.
Fig. 2: Example of the architecture of a typical opinion mining application used for a product review (Tsytsarau and
Palpanas 2012)
Noting however that these words were used in formal conversations, words would most likely
be used in informal conversations, at a personal level, in response to CCI publications or events
related to CCI were sought also. Words such as ‘tax’, ‘health’ and ‘income’ were identified
from comments on SMNs like LinkedIn, Facebook, Twitter, Reddit and the webpages of
mainstream news media houses; Bloomberg for example.
Identifying these words were one part of the challenge, understanding which words were likely
to be used in most conversations on SMNs was the other part of the challenge, no easy method
exists for. So we had to turn to other works in the literature, which had to devise some
9
nomenclature to mine data from SMNs. Then the syllable of these identified words were noted:
in private conversations single syllable words are aptly adopted than words with more than two
syllables. Since there were no one tools that could be used to justify this observation using data
from SMNs, another perspective was sought from other related work in the literature (Daas and
Puts 2014). A combination of these perspectives and consultation with supervisors led to the
selection of fourteen keywords used to in NetBase to extract data.
Research Contribution: There are claims in the literature that data from SMNs are noisy. In
order words their content are found be contain more irrelevant data than useful data, albeit for
the particular research purposes. Since there are no obvious attempts on how to quantify this
noise, one of the techniques used in this thesis to assess the representativeness of SMNs data
can be reused by researchers to gauge how well their mining tool, or keywords are performing
relative to a subject of interest. The method is simple: it uses the SMN with the highest number
of users, Facebook, as a benchmark. Then assuming the data mining tool aggregates summaries
like NetBase does, the volume of the most used word in the ‘top term’ is divided by the
benchmark, and multiplied by 100. If this number is very small < 1%, it would imply that a
small proportion of SMNs data are directly related to the subject of interest. This metric can
easily be calculated for each top word under the results obtained per keyword. Such that, if the
metric is small, it indicates that the keyword may not be adequate and so could be substituted
with another keyword. And the process can be repeated until this ratio becomes >1%, for each
of the keywords used to mine data. Thereby improving the representativeness of data got from
SMNs and in turn the accuracy of the information sought.
1.8. The structure of thesis It is necessary to understand the ongoing effort to extend the body of knowledge to the context
of this research study. Concepts have been postulated about the relevance of the CCI and the
significance of variables used in its computation. These concepts and research applications
published in the literature are reviewed in Chapter 2. Chapter 3 describes the methodology of
the research in-depth. It will cover aspects that have to do with the means through which the
end goal will be achieved. Chapter 4 deals with the processing of data to yield exploratory
insights into the richness of such data for the purposes of understanding patterns that relate to
the confidence of consumers in the economy and to know whether or not social media is a good
tool in measuring sentiments. In Chapter 5, the results of Chapter 4 are discussed to highlight
the learning outcomes; then a conclusion and recommendation for future research studies are
given.
10
Chapter 2
Literature Review
The use of data from social media networks to synthesize macro information continues to elicit
a tremendous amount of research interest. There is a huge wave of research applications
employing data from social media networks to compute informative metrics, such as the level
of customers’ satisfaction relative to quality of business service delivery. An example of this
is ratings that show the punctuality of train services based on customer’s complaints or
commendations on social media networks.
The first part of the literature review focuses on the concept of consumer confidence as an
economic indicator.
The second part concentrates on the tools and concept of data mining. This is important since
these are the interfaces through which data from social media networks could be extracted.
The third part of the literature review will explore whether data from social media networks
contains attributes that could give an exploratory perspective on variables that have been
identified as components of the consumer confidence index indicator.
2.1. The economic indicator of consumer confidence index
The publication of a Consumer Confidence Index started in 1946. It was based on the
University of Michigan’s Survey of Consumers. Since then, different aspects of its versions
have evolved. For example, today’s version includes derivatives like the frequency of its
measurement, its sample size and method of data collection, including the dependent variables
and independent variables, the rules governing whether respondents from a previous data
collection exercise could participate again, and what the maximum percentage of such
respondents would be are added (Merkle, Daniel M; Langer, Gary E; Sussman, Dalia;, 2004).
It has also been suggested that the CCI was established to produce an economic indicator that
would ‘lead’ imminent economic activity and thus be used to forecast consumer expenditure
(Roberts & Simon, 2001).
When it comes to how the CCI is determined, three of the oldest forms of measurement
constitute the foundation upon which today’s concepts are built. They are ABC News/Money,
11
the Conference Board (of the USA) and the University of Michigan consumer confidence
surveys. Pertinent information about their methods is shown in Fig.3 (Roberts & Simon, 2001).
All methods shown in Figure 3 are based on the assumption that sentiment can be used to
predict household spending, especially for durable products which are ‘discretionary’ in nature,
given the ease with which their purchase could be postponed. As Mueller (1963) stated,
“confidence variables were close predictors of durable and non-durable household
expenditures.” (pp. 899-917).
Methodology of Consumer Confidence Surveys
ABC News/Money Conference Board Univ. of Michigan
Method Telephone Mail Conference
Sampling RDD with random selection in
household
Selection from a non-
random panel
RDD with random
selection in household
Weighting For probability of selection
and to Census (region, age,
race, sex and education
Not disclosed For probability of
selection and to Census
(age and income)
Sample size About 1,000 (250 per week x
4 weeks)
About 2,500 for end-of-
month release; 3,500 for
later revision
250-300 for mid-month
release; 500 for end-of-
month revision
Field period Wed-Sun each week; Results
based on a four-week rolling
average
Sent first of the month;
Accepts returns through end
of month
Around first of the
month through a few
days before the release
Release Weekly, Tuesday evening Prelim. Figures, last
Tuesday of month; final
figures with next month’s
release
Preliminary figures at
mid-month; final figures
at end of the month
History Started in December 1985 Started bimonthly in 1967;
went to monthly in 1977
Started annually in 1946;
quarterly in 1952 and
monthly in 1978
Fig. 3: Differences between the three oldest forms of consumer confidence index measurement (Source: Reserve Bank
of Australia)
It has been noted that consumer survey results provide confidence indicators and important
information on assessment by each consumer concerning their economic situation and their
various expectations for the near future difficulties in application continue however. For
example according to Mueller, the Michigan Consumer confidence survey, an important
predictors of consumer expenditure on durable and non-durable goods was unable to ascertain
the future performance of the economy due to its lack of confidence variables at the time
(Mueller, 1963).
12
In 2006, Cotsomitis and Kwan demonstrated how they were able to forecast the expenditure of
households by combining two economic indicators. The first was the CCI which indicates the
economic prospects of the economy by measuring the level of consumers’ spending. The
second indicator is the Economic Sentiment indicator which identifies the incentive(s) behind
the consumers’ spending activities. This slightly differs from the approach of Katona (1960),
who claimed that the best way to measure the [economic] confidence of consumers is through
their discretionary expenditure.
Research has extended into components of the different methodological strategies used by
these three consumer confidence indexes. Chief among these components are the sample size,
the weightings used, the sampling and the method of data acquisition, such as telephone or
mail.
All of these components continue to drive the evolution of the consumer confidence index
which originates from consumer confidence surveys.
There is, however one fundamental limitation of survey based methods. It has to do with the
non-participation of respondents and this is deemed as a higher risk due to inaccuracy
compared to the risk of unrepresentative information that could result from the use of a small
sample size (Roberts & Simon, 2001). This is the case of the Conference Board of Canada. On
its website it is stated that the size of its sampled respondents is 2500 – 5000, but only about
1,500 respondents replied to the survey (The Conference Board, 2011) (Bloomberg News,
April 2003).
This limitation continues to spur research efforts to clarify and develop evidence on
associations that form the core of the economic indicator of consumer confidence. One such
study examines the association between the variables of sentiment and economic activity,
identified and discussed in the American literature. Nonetheless there is little evidence on the
strength of this association, why it exists and its causal direction (Roberts & Simon, 2001).
Yet in another regard, the index of consumer confidence is subjected to rigorous analysis,
especially when associations with economic activity are being discussed. Whereas some
national indexes of consumer confidence are derived by weighted or unweighted contributions
of other indices, which do not get a fraction of such critical analysis.
_____________________________________________________
2 (The Wall Street Journal, 2014) (FxPro Blog, 2014) (Reuters, 2014) 13
For example the University of Michigan’s CCI is determined using the contribution of the
index of Consumer Survey, a little known survey (Merkle, Daniel M; Langer, Gary E;
Sussman, Dalia;, 2004).
The majority of previous research focused on the economic indicators of consumer confidence,
and refers to evaluating how the different components used to determine it are used, ignored,
assumed to be significant or modeled. In particular, the choice of variables and their assignment
as being dependent or independent for the application of statistical procedures like regression
is considered. This continuous variation leads to a continuous revision of the consumer
confidence index even after it has been published2.
In summary, generating an index for consumer confidence is based on a method of data
collection that has not changed much since 1946. Instead, it is the methods of processing the
data, how the sample of the population that participates is rotated, the rules governing the
participation of previous participants in new surveys and the data models used to develop an
interpretation from the data that has gained widespread attention in the research community.
Considering the above, it is evident that the composite aspects of the consumer confidence
index, i.e. other indices like Consumer Sentiment used to determine the University of
Michigan’s consumer confidence indices, play a pivotal role in the indices representativeness.
Also, these other indices relate to aspects of human sentiment or emotion that could be traced
to and extracted from human interactions.
For example, in May 2014 the survey of consumers by Thomson Reuters and the University of
Michigan found that respondents expressed concerns about slow wage growth. This would be
indicative of a decline in living standards for households in the year ahead, after their income
is adjusted for inflation. In another context, respondents hinted at hesitant buying intentions,
citing fewer discounts on prices and higher interest rates (Thomson Reuters & University of
Michigan, 2014). This leads to the hypothesis that if people are happy to share the realities of
their economic situations in surveys, they are likely to do the same on social media networks.
In Fig. 4, the National Restaurant Association was able to survey its customers to understand
their personal finance and they were willing to share this information. The outcome was
described by the following statement, “Nearly six in 10 adults say their household finances are
either in fair or poor condition, which is essentially unchanged from their appraisal of
14
their finances in 2010, when the economy was in the early stages of recovery” (National
Restaurant Association, 2014).
Fig. 4: A survey of restaurant customers on their assessment of their personal finances
Clearly, this presents an opportunity to use sources of information where human interaction is
the exchange currency—and this source could potentially be social media networks. Since the
broad goal of this research is exploratory, it is worth attempting to show that data from social
media networks do have and reflect at least a few representative similarities with the
information obtainable from composite indices of the CCI.
But to do so will require a look at what transpired from the body of published knowledge that
continues to shape how associative models are used to link human interactions with various
subjects of interest. The body of knowledge that has propelled this interface of information
extraction will be reviewed next in subsection 2.2.
For completeness, a review summary is given in Table 1, where consumer confidence indices
from around the world are explained in terms of their methodology and approaches.
15
Table 1: Summary of Consumer Confidence Indexes from around the world
Canada United States of America United Kingdom France Nigeria
Organization
Responsible
Conference Board Conference Board and
university of Michigan
consumer sentiment Index
GFK and MORI. Commission in the Journal
European Economy
(Supplement B).
MasterCard worldwide index.
History The conference board started
its operation in 1980 and has
been an independent research
organization since inception.
(Conferenceboardcanada
2014).
The Conference board
started in 1967 and
Michigan’s Consumer index
started in 1946.
GFK started across Europe as
early as 1970 but officially
started in UK in the year
1995.
Method Telephone The conference board uses
mail to acquire its own
consumer confidence data
while the Michigan uses
Telephone as its method of
acquiring the data.
Telephone. Gathering information based
on usage of master cards by
the consumer.
Size Approximately 2,000
households are used for the
survey.
The survey is carried out on
approximately 2000
households.
Timing and Data
Release
In 1980 the consumer
confidence data was produced
once in four months, in 2002
the pattern was changed and
the data was then produced
monthly.
Both organizations starts
carrying out their survey
first day of each month. The
conference board releases its
result at the end of the
month whereas Michigan
publishes its initial result on
the second Friday of the
month which is around 50 –
60% of the total number of
response collected. The final
results are published at the
end of the month (Curtin
2003).
The survey is carried out and
published once every month.
GFK focuses its questions on
household finances and the
general economic outlook.
This telephone numbers are
selected randomly and dialed
every first three weeks of the
month except in August. July
and September are used to
calculate August.
This survey is carried out
twice every year in the various
geographical region in the
country. (Vanguardngr 2011).
16
Operations and
Questions
The consumer confidence
survey is bases on four specific
questions:
Considering everything,
would you say that your
family is better or worse
off financially than 6
months ago?
Again, consider everything,
do you think that your
family will be better off,
the same or worse off
financially 6 months from
now?
How do you feel the job
situation and overall
employment will be in this
community 6 months from
now?
Do you think that right
now is a good or bad time
for the average person to
make major outlay for
items such as a home, car
or other major item?
The indices produced by
both Michigan and
Consumer board based on
questions on the current
condition of the economy
and questions about future
expectations.
For the measurement of the
current condition the
consumer board questions
are only directed towards
the view of economic
conditions and not the
personal experience of the
consumer whereas Michigan
questions are directed
towards family finances and
the peoples spending in
general.
In order to effectively
measure future expectations
both the conference board
and Michigan assign two
question directed
specifically towards the
economic condition and one
question directed towards
the consumer’s personal
finances.
One of the major differences
is the structure of the
question, Consumer board
tends to ask questions about
the consumers view of the
condition in their local area
but Michigan aims at
gathering data based on the
customer views about the
condition of the nation as a
whole rather than their local
environment.
The approach of questions used
by these bodies was widely
criticized by Dominitz and
Manski (2004).
GFK survey in the United
Kingdom is based on five
specific questions. They
include:
How has the financial
situation of your
household changed over
the last twelve months?
How do you expect the
financial position of your
household to change over
the next twelve months?
How do you think the
general economic situation
in this country has
changed over the last
twelve months?
How do you expect the
general economic situation
in this country to develop
over the next twelve
months?
In view of the general
economic situation, do you
think now is the right time
for people to make major
purchases such as furniture
or electrical goods?
The questions asked are
based on five indicators, they
include:
Personal financial
position –Past change
Personal financial
position- outlook
Likelihood of major
purchases
Living standards in
France- past change
Living standards in
France- outlook
MasterCard index worldwide
survey questions in Nigeria
are based on the following
sectors
The Economy as a whole
The rate of employment in
the country
The stock market
Regular income of the
consumer
Quality of life experience
or expected by the
consumer
17
Computation of
Indices (Positive
and Negative)
After collection of consumer
confidence data, the positive
and negatives are used to
compute and analyze the data
in order to provide a result to
the general public on the
consumer confidence of the
country’s economy. When the
positive and negative has been
sorted out the index is then
computed for each question
using the formula:
Percentage of positive
responses/ (percentage of
positive responses +
percentage of negative
response)
According to Daniel M.,
Gary E., Dalia S.,
“The Conference Board
index is computed by taking
the positive percentage for
each question divided by the
sum of the positive and
negative percentages. This
number is then divided by
the base year value from
1985 and multiplied by 100.
The resulting values from
each question are averaged
to form the overall index”
“Michigan computes its
index by taking the
difference between the
positive and negative
percentages for each
question and then adding
100 to each. These are
summed and then divided by
a factor representing the
base year, 1966. Finally a
small correction is made to
account for a design change
in the 1950s”.
The indices are computed
using the positive and
negative values. According to
Stuart B., Melissa D;
“The results are published as
a net balance of positive less
negative responses, with those
who answered ‘a lot’ in either
direction given twice the
weight of those who answered
‘a little’”.
The consumer confidence is
the calculated using positive
and negative responses with
formula;
% of positive response - % of
negative responses.
The index is computed using
positive, negative and neutral
value. 0 values is assigned as
the negative value which is
known as the most
pessimistic, 100 is seen as the
most positive value which as
represented as the most
pessimistic and 50 is seen as
the neutral value.
18
2.3. Measurement of Consumer Confidence
The management consultancy firm McKinsey, published a report entitled “What executives think
about the economy (2014)” – that aimed at deducting the confidence of business executives relative
to the confidence reflected by consumers in their respective geographic business locations globally
(McKinsey&Company,2012).
Nielsen Global Group is another organization that measures consumer confidence since 2005.
From across Africa, Asia, Europe, Middle East, North America and Latin America 30,000
respondents are surveyed in order to quantify their economic sentiment/outlook in the ‘economic’
segments of their lives.
The Nielsen Group report revealed that the global CCS data collated in the last quarter of 2013
shows that CCI has remained the same with an index of 94 points (Fig. 6) for three preceding
quarters unlike the year before which was at a 91 index point. This means that the global consumer
confidence has increased since 2012.
Although this CCI figure is for all the countries involved, an increase in the global index data is
not representative of an increase in respective individual countries. The last survey carried out on
the 29th November 2013 was the fourth quarter in 2013, and it showed a CC increase of 43% of
the markets which was lower compared to the 57% increase which was measured in the third
quarter of 2013. The highest consumer increase was seen in Indonesia with a point of 124 whereas
the lowest decrease was seen in Portugal, Italy, Croatia and Slovenia with a declining index point
of 44 (Nielsen, 2013).
We can see that the low CC of Portugal and Italy, for example, correlates to the findings of the
World Economic Situation and Prospects report (United Nations 2015), that indicated the
weakening of these economies owing to a slowdown of export driven business in 2011 and 2012
(United Nations, 2015). In turn this slow down impacted the profitability of local businesses that
enjoy significant export business, thus leading to less capital investment, cost cutting measures
and other cash flow sustenance measures that further impact the economic situation of the
workforce who is also the primary consumer source. Thus it is these impacts that Nielsen has
managed to elicit through its surveys.
19
In Fig. 5, a hand count of the number of countries that experienced an increase in CC shows there
are 26 compared to the 32 with a decreased level of CC and 2 countries that experienced no change.
As such it is evident that the CCI measured in the preceding three quarters, i.e. January to October
2013 must be a net increased CC around the world.
Fig. 5: A global CCI of sorts generated from a survey of 60 countries for the period of October – December 2013; Source:
Nielsen, Consumer Confidence Concerns and Spending Intentions around the World Quarter 4, 2013.
2.4. Social media networks
Social Media is composed of online technology tools which enable communication between
different people from around the world via the internet.
According to Margaret Rouse 2011,
“Social media is the collective of online communications channels dedicated to
community-based input, interaction, content-sharing and collaboration.”
20
For use in the economic and business sector, Social media involves the gathering of comments
made in online forums, created with the sole purpose of educating people about products, brands,
services and general issues in the economy (Blackshaw & Nazzaro, 2004).
Magmold and Faulds (2009) defined social media as a wide range of word-of-mouth forums which
allows users to be able to express their feeling about a topic or even create a topic for discussion.
The forums included: social networking sites, blogs for personal and company discussions and
product rating sites.
There are different types of social media in the world today, they include for instance, Facebook,
Twitter and LinkedIn. The communicating parties are able to connect and hold conversations using
formats such as; text, audio, video, images and other multimedia.
Social media has become one of the strongest and most powerful tools used online since it allows
thousands of users to be reached within the shortest minimum time possible. According to statistics
provided by socialnomics.net, on the speed of development of the social media, it took the radio
38 years to reach 50 million users, TV 13 years to reach 50 million users, but it took Facebook just
9 months to reach 100 million users.
The tremendous difference in outreach suggests that social media is an extremely fast way to get
to customers or users about a particular product or issue. Social media has not only been used for
personal use but it has been used for business as well. Using social media has become one of the
new methods used by The Economist to gather data on the level of consumer confidence in the
economy, the methods including: counting the number of likes and dislikes of a topic, how many
comments have been given by consumers on relating topics and gathering how many re-tweets
were done by consumers.
This method of gathering data shows the real consumer sentiments towards a brand or the economy
based on the comment provided by the consumer. According to a report from Performics, 2010, a
survey carried out on 3,000 U.S social media users revealed that social media has a tremendous
effect on customers attitude towards brands. The report also showed that 53% of the authenticated
Twitter users recommended a company and its products by tweeting about it and 48% of the people
actually delivered on their intentions to buy that particular product (Performics, 2010). According
21
to Keller Fay Group in 2010, the average consumer in a country mentions his/her specific favorite
brand approximately 60 times in a conversation with family, friends and on social media.
SMNs are rich with expressions of sentiment which is important to the measurement of CCI.
Sentiment is known as the feeling or reason for an expression made behind a comment and
reference about a particular topic or brand. Various tools have been made which enable measuring
the consumers’ sentiment. From a business perspective, the ability of a business owner or a brand
to be able to measure or understand the consumer’s sentiment behind a comment is very beneficial
to the growth of the organization. This helps to determine the state of mind a consumer was in as
at the time he/she made the comment.
Research was carried out on how to effectively predict movie sales from bloggers sentiments
(Mishne & Glance, 2006). The researchers stated that the volume of discussion about a product on
the social media especially web blogs, cannot guarantee a product’s final performance, therefore
they advised that sentiments on the weblog data should be analyzed as well as the volume of the
data, concluding that positive sentiment - which is different from positive result- is actually a better
data for the prediction of the success of the movie, rather than just the positive comments.
A similar analysis was carried out on how social media can help predict the future outcome of a
new movie before it is released. In this research, Twitter.com was used to forecast box-office
revenues for movies instead of the static blogs which were used by Mishne and Glance. The result
from the research was calculated by analyzing the number of tweets that are currently on the
network based on a particular movie.
These tweets were divided into positive and negative tweets, Positive tweets are interpreted as
representative of customers looking forward to watching the movie based on the preview, while
the negative tweets mean that the customers are not interested in watching the movie. A linear
regression model for predicting box-office revenues of movies in advance of their release was
constructed. The results showed that there is a strong correlation between the amount of attention
or discussions consumers have about a movie about to be released and its subsequent ranking when
the movie is eventually released (Asur & Huberman, 2010).
Looking at the two research studies it can be concluded that the result derived from the dynamic
method (tweets) can be used by itself to forecast the revenue to be made from the forthcoming
22
movie whereas the result derived from the static method is not enough to predict movie sales, the
sentiments behind the content of the blog must also be considered.
Chris Barry, Rob Markey, Eric Almquist and Chris Brahm (2011) used a Net Promoter system to
measure how the social media has affected the consumers’ confidence on a product or the
organization as a whole. The questions asked on the social media about the product included “How
likely would you recommend [this company or product] to a friend or colleague in social media?”
After analysis was carried out, they stated that:
“Customers who engage with companies over social media are more loyal and they spend
up to 40 percent more with those companies than other customers.”
In 2012, Schweidel, Moe, and Boudreaux carried out an analysis of the potential to get brand
sentiment from social media conversations. They used data which was collected from different
social media domains such as Facebook and Twitter. They proposed the use of a hierarchical
Bayesian regression model which was used to measure these sentiments effectively.
Based on the research carried out, it has been discovered that social media is one of the fastest and
easiest mediums that can be used to get large amounts of consumers reviews which can be used
eventually to calculate the overall consumer confidence on a brand or the economy as a whole.
Different approaches on what actually affects consumer confidence were also researched. Among
them are unemployment, rising inflation and stock price, but the greatest question still remains
“What actually causes a rise and fall in consumer confidence?”
In conclusion, the social media itself can actually cause a rise and fall in consumer confidence in
the world today as information about the economy can easily be known by consumers as quickly
as it happens (using blogs, Facebook, twitter). This affects how consumers feel about the economy
unlike the days when such platform was not easily accessible.
2.5. Attributes of Social Media Networks
Social media in recent years is large and well known for its social network and content sharing
abilities—videos and pictures (Asur & Huberman, 2010). It can also be seen as a setting for virtual
discourse where people create content, share it, bookmark it and network at an extraordinary
23
degree and speed. It is also used as a real-time snapshot of updates of interest, location, instant
memories and so on.
The use of social media networks has brought about one of the biggest cultural shifts since the
industrial revolution, attracting up to 1.82 billion users globally (Statista, 2014). It is engendering
a whole new wave of human interaction, different from the structured methods that existed before
it. The increased use of mobile internet technology has also contributed to the growth in adoption
of SMNs illustrated in Fig. 6. Some examples of popular SMNs as seen in the figure include:
Facebook, Twitter, MySpace, Digg, Reddit, Instagram, foursquare and sound cloud.
Social media comes in various forms:
Wikis: This act is done by adding articles and editing existing articles. Example;
Wikipedia, Wikia.
Social News: This act is done by voting for articles and commenting on them. Example;
Digg, Propeller, Reddit.
Social Networking: This act is done by adding friends, commenting on profiles, joining
groups and having discussions. An example would be Facebook, Hi5, Last.FM
Social Photo and Video Sharing: This act is done by sharing photos or videos and
commenting on user submissions. Example; YouTube, Flickr.
Recruitment: This act is done adding suggestive jobs available, commenting on groups
formed by companies, sharing various job experience and also applying for jobs. It is the
main source for e-commerce recruitment. Example; LinkedIn.
Social Bookmarking: Example is Del.icio.us, Blinklist, Simpy. This is used by tagging
websites and searching through websites bookmarked by other people.
24
Fig. 6: Global Penetration of social platform-Account ownership and Active usage— (Chaffey, Dave, 2014)
2.6. Data mining concepts
The mainstream approach of generating the economic indicator of consumer confidence is through
surveys and it has proven its reliability over time (Merkle, et al., 2004). Nonetheless the process
of generating this index is becoming hybrid, where traditional survey methods of mail or telephone
calls are used to gather data and the information extraction models are used to extract additional
data from other [electronic] sources. This is a new approach that is gathering momentum,
especially the use of social media networks. The Conference Board of the U.S recently outsourced
the creation of its consumer confidence index product to Nielsen who uses this hybrid approach
(The Conference Board, 2011).
Also, there is an uptake in the use of data from social media networks for the generation of insights
into multiple areas of interest related to human interaction. At the core of this process are models
that aim to associate the textual content of social media networks with the intentions of the
25
interacting parties. This has driven research efforts in the area of information retrieval, commonly
known as IR.
Information Retrieval (IR) methods have been developed to optimize the extraction of insights
from data in social media networks. NLP and LIWC are the well-known examples. In the financial
services industry for instance, innovative companies such as Barclays, TD Canada Trust and the
Bank of America leverage such data to succeed in very difficult sectors like pay day loans, a
banking business segment whose clients are known to default often.
Wonga® a payday loan company in the UK, pioneered the use of data from social media networks
to mitigate the industry-wide challenge of defaults. It uses up to 8000 data points automatically
captured from social media networks to augment the profile of borrowers based on information
they have provided. This data is then aggregated to classify whether the prospective borrower will
default or not. It is claimed that this application of information from social media networks reduced
the percentage of defaults by 50%—a record breaking feat going by the industry’s standards
(Tobin, 2012).
In light of the fact that the storage of human communications is in text format, models of lingual
origin continue to be developed to correctly associate meaning, context and intention expressed in
such pieces of information. In fact this remains the challenge plaguing various frameworks that
have been proposed in the body of literature.
It turns out that in order to retrieve information from a textual source, word content must be
linguistically classified, after which they are associated to a segment of human life (e.g. emotion,
religion, social, economy and cultural). Then they are aggregated to piece together the intent,
context or meaning of the interaction held by the ‘writer(s)’ and the ‘listener(s)’, as it may be. The
reality is that there is no one framework that captures all three correctly.
Frameworks have therefore been targeted at either deriving meaning or the intention or context of
a communication made known on social media networks in particular. A good example is a
framework that uses argumentative filtering to establish whether a piece of information expresses
a negative or positive point of view/intent. Other frameworks apply concepts such as NLP or
LIWC. This research study simply aims to investigate if the attributes of data extracted using such
26
micro level models is good enough to synthesize macro information such as the generation of a
CCI.
Although there is a high proportion of visual-audio content in some social media networks such as
YouTube™, these are excluded from the scope of this research. Based on observation, social media
interactions predominantly involve text based communication. The majority of users would much
rather write a response or comment, instead of posting a video or audio response. This is the case
when matters such as government policy, labor rates, unemployment, tax increases or the economy
in general are being discussed, in contrast with the visual media which is more of a social
interaction tool for marketing and other forms of ‘soft’ topical interactions. Of course, most
executives, organizations and governments post videos to engineer engagement when discussing
matters of a serious nature. But this is a small number compared to the majority of social media
network users, who also play the role of consumers in the economy.
There are a variety of concepts that postulate models for mining textual content for information.
In some instances they have been used to predict the sales of movies or even predict the outcome
of elections. The U.S presidential election won by President Barack Obama in 2008 is a well-
known example of predicting wins (INSEAD, 2009).
2.6.1. Natural Language Processing concept
NLP is a branch of computer science that deals with the processing of human language so that it
is accessible to computer applications, in such a way that it enables applications to process,
classify, aggregate and generate insightful information of a higher human understanding
(Microsoft Research, 2014). The earliest efforts of NLP application started in the 1950s and since
then NLP has been behind a diverse number of applications. Examples of such applications
include: information extraction, machine translation, sentiment analysis and question answering
(Research at Google, Google Inc., 2014).
These have been materialized into products like the voice recognition applications for smart
phones and cars used to identify the task their user intends to perform, spam detection in electronic
email, the Online web chat used by companies for web based customer support services, the IBM
Watson computer and so many more (Jurafsky & Manning, 2014) (Research at Google, Google
Inc., 2014) (IBM, 2014).
27
NLP is a fusion of a long list of tasks that solve a piece of the language processing puzzle, as the
incentive to reduce the latency of language translation for computer processing grows strong.
Some of these tasks have led to solutions of a basic level of language processing. For example,
Name Entity Resolution which has to do with identifying words in a text that refers to people,
places or institutions, and Part-of-Speech (POS) tagging where a word in a text is identified as
being either a noun, verb or pronoun (Collobert, et al., 2011).
NLP uses semantic rules at different levels to associate components of language. For example, it
analyzes nouns and phrases, tagging each to label it as either a person or organization and
clustering a collection of these components so that those that refer to the same entity are easily
classifiable.
The depth of this breakdown of language based on rules depends on whether the source of the data
is closed or open and unstructured like the internet, Web 2.0 (Research at Google, Google Inc.,
2014). Historically, most NLP research has focused on the former and only few companies like
Google are advancing the state-of-the-art of NLP, where the text being processed is varied and
does not conform to the structure of spoken language.
Similar to the fundamental constructs of language, NLP employs the syntactic rules of structured
language to predict part-of-speech tags for each word contained in a sentence or document being
analyzed. This sort of ‘parsing’ has proven to be a fundamental challenge for modern day
applications of NLP, which have to deal with multiple human languages without any communal
lingual structure.
Other tasks are aimed at tackling language processing challenges that are at an intermediate level
and steady progress is being made. Examples of these include Sentiment Analysis where the goal
is to extract subjective information from a sentence or body of text and thus determine its
“polarity”. Polarity is an indication of whether it is for example positive, negative, in-support and
against, depending on the use case objective. Others include Information Extraction, e.g. the
extraction of a calendar date from an email invitation specifying the date and time of an event.
Then there is Word Sense Disambiguation (WSD) and Machine Translation to name a few
(Jurafsky & Manning, 2014).
28
Finally there are tasks that have to do with macro information synthesis. These border on the
production of insights that are deeper than the lateral interpretation of language. For example, in
the summarization section we have the following sentences (i) the London Exchange is up (ii) the
Standards & Poor 1000 index rose and (iii) employment jumped, the summary that could be
deduced is that the economy is good. This is a herculean task for a computer to undertake, because
it needs more than just semantic and syntactic rules of language to reach such insight. It needs
knowledge and experience equivalent to that of an individual that is aware of the existence of S&P,
the London Exchange and the fact that they are strongly associated to the state of an economy.
Other tasks in this category are Paraphrasing, Question Answering (QA) and Dialog for web help
services (Kongthon, et al., 2009) (Jurafsky & Manning, 2014).
NLP is a concept that has enjoyed considerable research and advancement. An aspect of NLP that
could be applied for this research is Sentiment Analysis and there are other aspects such as Parsing,
Summarization, POS and WSD.
Assuming that the majority of the data on social media networks follow language rules, NLP is
one approach that would enable the extraction of relevant classes of conversations, particular
words and possibly the polarity of conversations relative to the economic situation of the
conversationalists. Therefore, the selection of a tool that would be used for this research would be
anchored on whether it implements NLP algorithmically.
2.6.2. The Linguistic Inquiry and Word Count concept
The Linguistic Inquiry and Word Count (LIWC) methodology, developed by Pennebaker, Booth
and Francis in 2007, was designed specifically for analyzing text in order to measure the emotional
expression it contains. Since this research is being undertaken at a macro level, the LIWC concept
is not the right tool. Another approach is the Life History Calendar. This concept is based on
cyclical occurrences in life, such as birthdays, weddings, childbirths among others. It is believed
that an expression of one’s disposition towards these events would give an indication of their
situation at that point in time.
2.6.3. Life History Calendar (LHC)
It is a concept that is used retrospectively on data about events occurring in people’s lives. Such
as: when they got married, graduation from or enrolment into a new education level and first job.
29
In 1969 a national sample study was carried out in the United States using the LHC technique. In
this study, various kinds of data such as: timing of education, employment and family events were
recorded through the use of the technique (Freedman et al, 2011). The LHC concept is biased
towards seasonality, an attribute which constitutes a small part of conversations on social media
networks. For this reason it does not represent a useful approach for this study.
LHC has two basic components:
A visual feedback of answers which includes multiple life domains; examples of data in
this category are work, education, health and household composition.
A landmark event. Examples are public and personal.
A.2.1.1. Advantages of Using LHC Technique
Improve the quality of the retrospective: The use of the LHC concept ensures that events
relating to an individual is stored and gathered; therefore the quality of the retrospective data
gathered can be improved as the respondent is able to relate both with the timings of the events
visually and mentally. This respondent can also use the full patterns of their life events which
were recorded to remember the timings of past events accurately (Perl 2002).
The nature of LHC enables gaps in the data to be detected thereby Improving completeness
of reports.
The calendar format and the visual cues help interviewers collect complex life history data by
structuring questions within timing and sequencing framework, while flexible recording
techniques allow interviewers to adjust their questions in accordance with respondents’ ability
to recall events (Perl 2002).
2.6.4. Method Proposed by Nigam and Hurst
In order to determine or detect expressions behind a topic or comment made by an individual,
Nigam and Hurst (2004) proposed a method that involves two distinct components. These are:
Shallow Natural Language Processing Polar Language Extraction System (SNLPPLES).
Machine Learning Based Topic Classifier (MLBTC).
These components work hand in hand to aid in making a simple and accurate assumption which
states that if a topical sentence contains polar language, the system predicts that the polar language
is reflective of the topic, and not some other subject matter (Nigam & Hurst, 2004). In this method,
30
the polar language is first identified in the individual sentences by the system, the machine learning
based Topic classifier is then trained with machine learning techniques using a collection of
documents which relate the topic, the document classifier is then adapted into a high precision/low
recall sentence classifier.
During the analysis, if the classifier judges the document as irrelevant, then all the sentences in the
document are assumed to be irrelevant as well; but if the classifier judges the document as topic-
relevant within the topic specified, then the sentences are assumed to be relevant, further
examinations being then carried out on the entire sentence in the document. Nigam and Hurst
(2004) went further to use the Bayesian statistics to form an aggregate authorial opinion metric.
This metric was used to propagate the uncertainties which could be introduced by polarity in other
to statistically facilitate valid comparisons of the various opinions across multiple topics. The
Nigam and Hurst theory was used by Mishne and Glance (2005) to predict movie sales from
blogger sentiment.
The concept proposed by Mishne and Glance (2005) is judged unsuitable for the purposes of this
research, because it relies on file based content, and that does not apply readily to data from social
media networks.
The concepts that have been reviewed thus far would be applicable if the goals of this research
were non-exploratory. In other words, if the goals of the research were to generate a consumer
confidence index based on a single source of data the above approach would add value. In spite of
this, knowing that there are a variety of concepts that could be used to extract information. In the
next section, the main tool chosen for this study to extract data from social media networks will
be reviewed.
2.7. Data mining tool
2.7.1. About NetBase®
NetBase is capable of extracting data which contains both an aggregative indication of occurrence
(across random demographics) and raw natural language text posted on different social media
networks. NetBase implements the concept of Natural Language Processing (NLP) and it uses it
31
to determine the bias of a comment, opinion and even emotions expressed on social media
networks (NetBase, 2014).
The NLP is very effective as it does not count the words or analyze word for word like a
conventional method such as LIWC. Instead, NLP uses the entire sentence, then the patterns of the
grammar/sentence are analyzed before being organized (GNIP, 2010). It uses two specific steps
when analyzing the data:
Parsing: This is the first phase in the analyzing stage. In this stage the NetBase NLP engine
parses each of the sentences gathered from the social media. It then identifies the various
subjects, objects, verbs, adjectives and other patterns in the sentences at the analyst‘s
discretion, thus enabling a deep and accurate understanding of the comment to be extracted
(GNIP, 2010).
Normalization: This involves defining a standard or an algorithm for how the sentence would
be analyzed. At this phase the NLP engine defines the appropriate algorithm for matching
keywords thereby preventing inaccuracy (GNIP, 2010).
2.7.2. Advantages of Using NetBase®
Accuracy: NetBase uses a sophisticated Natural Language Processing (NLP) engine for
gathering and analyzing the data (Netbase, 2012). This engine ensures accuracy in the data
collected. According to a report, NetBase is nearly 90% accurate (GNIP, 2010).
Speed: NetBase guarantee real-time data and analyzes as it is carried out on social media
around the world.
Ease of Use: It is very easy to use. No required experience is expected to be able to read the
data as the software performs the gathering and analyzes. It also groups the comments as
positive, negative and neutral thereby allowing the analyst to be able to measure the impact of
their brand or organization across the various social channels.
2.7.3. Limitations of NetBase®
License is Expensive: The license key for the software is quite expensive and may not favor
small and medium enterprises. Only the large organizations are able to purchase this license
as they feel they are at greater loss and they should be aware of what their consumers think
about their brands or products.
32
Extracting and Filtering of Data is left at the analyst’s discretion: Not all the data is extracted
and filtered. The analyst has the power to determine what should be filtered and extracted, and
therefore there is a possibility that they can be biased on their decisions.
There are other data mining tools that implement a concept of data mining mentioned or a
derivative of one. The limitation of some of these is that they are specialized and optimally
designed for applications with a narrower scope compared to this research. As a result, NetBase®
is preferred in the present research because its designed scope of utilization is broad as it
implements NLP.
2.8. Chapter Summary Some awareness of the economic indicator of Consumer Confidence indices has been attained,
leading the way to explore how it has been measured historically and how the methods of
measurement have evolved up until now.
In an age of social media networks, an opportunity exists to explore this new data source whose
content is mainly text based but increasingly encompassing other forms of conversations. This is
also the objective of this research—to explore whether this data source is sufficient to synthesize
information from a diverse dataset of uncommon language rules, syntax, semantics and structure.
The review of the literature in the field of data mining has shown there is a strong interest in the
application of social media networks and their data for analysis of real world problems. As such,
in order to harness this data source, concepts such as NLP that have been in the fore of research
interest since the 1950s were reviewed and an introduction given to other data mining concepts.
Other advances that have been made recently are in the area of interpretation and association of
the content of social media networks with the intentions of the participants. An organization such
as NetBase has done well in their application of the concept of Natural Language Processing for
information retrieval. By choice it is their user interface tool that was used to extract data from
social media networks for the purpose of this research study. In the next section, the
methodological implementation of this research is presented.
33
Chapter 3
Methodology
For this research, social media networks such as Twitter, Facebook, Blogger, Google+ and
Instagram constitute the sources of the data collection. Each of these are categorized differently
and present a different perspective of human interaction. For example, Facebook allows the posting
of a mood status—‘what is on your mind’; Twitter encourages the use of 140 letter characters to
tweet ones viewpoint on a subject of interest and Blogger allows people to present their viewpoints
as a story.
For the current research, a dataset of 500 observation was used. The raw dataset of 500 records
was stored alongside the summarization of the data by NetBase® relative to the search term. Each
raw dataset contains metadata on the location where it originated from, the source (a website like
Facebook, Twitter, and blogs), date, time of the day it was posted and the actual text of a posted
communication on any one social media network.
3.1. Social media data collection
Social media networks in which the content is in text format are the sources from which data will
be extracted using the NetBase® data collection tool. The tool extracts Tweets, Facebook status,
FourSquare check-ins and blogs it associates with particular search terms.
The type of data collection procedure is search driven because the primary data collection is
automatically executed through electronic algorithms built into the NetBase®. Such algorithms
implement the aspects of NLP in order to provide data related to the subject of interest, i.e. the
keyword. (The keyword is implicitly related to the subjective disposition of users about their
economic situation). The tool selected for this research is NetBase®, already introduced in the
review of data mining tools.
Since knowledge of the model assumptions used in NetBase’s NLP algorithm is unknown, the
research used specific [search] terms to extract data deemed to be associated with economic
sentiment expression for the purposes of this research. Each search is done for a time period
34
covering 3 months (Jul. 2013 – Oct. 2013), 6 months (Apr. 2013 – Oct. 2013) and 1year (Oct. 2013
– Oct. 2013).
The so-called “search terms” are the result of extensive research into what constitutes the variables
used in the measurement of the consumer confidence index as an economic matter. The choice of
these terms stems from their association with other economic terminologies and likelihood of them
being used within social discussion circles. A good example is the use of employment, as it relates
to the total number of employed persons and the unemployment rate (Roberts & Simon, 2001) .
3.2. Keywords for search
Based on the developed attributes on economic sentiment expression the following terms were
used to search NetBase® for data.
1. Capital Expenditures— this term is highly used in corporate communication and news
media and it has a strong link to investment in physical assets by governments, businesses
and individuals with significant asset footprint—machinery or real estate
2. Capital Investment— is an investment or expenditure that is expected to generate future
economic benefits. It can also be seen as a long-life asset— (Investopedia, 2014).
3. Consumer Confidence— this refers to the level of optimism, trust, believe and perception
of consumers in the economy. It is an economic indicator that helps measure the mentioned
attributes above. According to The Wall Street Journal (2015), it states that the Conference
Board had a projection for January 2015, “consumer feel more optimistic about job
prospects and future labor markets” – (Madigan, Kathleen ;, 2015)
4. Employment Condition—we refer to the level of unemployment and employment rate
5. Family Prospect— more likely to be used in informal discussions is the term ‘family’.
However, there is a reliance on NetBase® to augment both words in order to extend the
catchment of inferred circumstances for ‘prospect’. When we say ‘family prospect’ it is a
future outlook of a family’s economic prospects, whether they afford what they want or
not. Example; if a family has an optimistic view or projection on better job, it equals to
more shopping and more holidays.
35
6. Financial Security— here we mean shares, bonds and derivatives. All these are resources
that can be used to support living standards both now and for future needs. It also means
being
7. Health— given the ubiquity of expressions of illness or of wellness, there is a direct effect
of a person’s health situation on their financial cash flow. At a higher level, a general talk
of the health of a person, an organization, government or community is also indicative of
the economic situation of the subject of interest.
8. Income Sources— this can be referred to as various sources of earnings that accrued to
consumers.
9. Liabilities— represent claims against the asset of an entity. This can also refer to
borrowings from a financial institution or person, for the development of one’s economic
status that is usually payable with a short term or long term plan.
10. Political Affiliation— Refers to the relationship that exists between members of a group
that seek to achieve common goals, ideas and morals through exercise and gaining of
political influence.
11. Recurrent Expenditure— this is a term meaning the ongoing expenses made by the
Government and other organizations. It could be salaries, wages of workers, scheduled bills
such as mortgages and loan payments. Capital assets such as bonds and stocks are not seen
as recurrent expenditure.
12. Retirement— here we mean people that have retired from active service. After a person’s
service years with no pension plan or retirement benefit system available, the outlook on
that person’s economic perspective or opinion will be affected.
13. Social Status— refers to various social classes, from elites to masses. A person’s social
status can be earned through different means: own achievements, inherited positions and
primarily by identifying one’s character. The primary social character sometimes shapes
ones entire life and defines their status in the economy. “One’s status tends to vary with
social context. For example, the position of a man in his kin group helps determine his
position in the larger community”- (Encyclopædia Britannica, 2015).
14. Tax— this is a compulsory monetary contribution that is levied by the Government on
business profits, employment income, goods and services. The word tax can be ubiquitous.
It has a direct and indirect influence on an economy. Example: if the tax percent goes up
36
in an economy, it can leave the people in a pessimistic state regarding how they view the
given economy as a whole.
These fourteen associative terms are deemed as ‘dimensions’ of consumers’ (economic)
confidence and they originated from search & extract from social media networks, data that could
be used to project an estimation of consumer confidence for the purpose of the study. The selection
of these dimensions was based on insights gained by the researcher through a critical review of
concepts.
A limitation of this approach is that human communication in text does not always contain key
words about the subject of the discussion, its context or the intention of the discussion. Therefore,
search results are bound to inaccurately omit posts or blogs or tweets that are classed as having no
correlation to the context, meaning or intention of the searched word key word.
For example, a status post could indicate that a user has expressed they are sick/ill/unwell, which
would require an extraction tool to discern whether that user is having difficulty with their bills or
job as a result. In order to contain this limitation, multiple search terms were used to extract data
from social media networks using NetBase®.
Each search term is objective by selected by its association with the expression of psychological
financial sentiment, economic situation news, opinion on economic events and life events that have
a bearing on economic circumstance and vice versa.
3.3. Research data analysis method
The extracted data was analyzed to explore its attributes and determine whether it is good enough
to measure CCI. The data from SMNs were analyzed in four ways to achieve this goal. The first
part of the analysis was qualitative: Part A. It focused on the infographic summaries of SMNs data,
aggregated by NetBase® automatically. In the second part of the analysis, Part B, the numbers
NetBase used to arrive at the infographic summaries were analyzed. This was a quantitative
analysis that looked into obscure but important ratios. For example taking the volume of the most
used word for a particular keyword and dividing it by the number of users of the largest SMN,
Facebook, was used to test for representativeness from a quantitative perspective.
In the third part of the analysis, Part C, the interest moved to the contribution of SMNs to the data
content extracted using NetBase. It was used to identify the SMNs with most contribution and also
37
to see whether the most popular SMN, Facebook, would also be the largest contributor. Then a
fourth type of analysis, Part D, focused on the raw data of the actual conversations from SMNs. It
looked into the classification of a selection of raw Soundbites to evaluate whether NetBase’s
classifications are correct. This analysis was also used to identify flaws in the configuration of the
data extraction mechanism of NetBase. In particular whether the same Soundbites were collected
and so repeated, if they were posted by a single user.
Finally, a statistical analytical tool, KH Coder was used to examine fundamental performance of
the NetBase® data mining tool and the NLP model it implements. Especially on its ability to
comply with basic language semantic rules, such as the occurrence of words that would normally
co-occur in spoken or written format. This fifth analysis was used to gain confidence on the
performance of the NetBase® tool at a fundamental level.
The reader is reminded that the premise of this research is fundamentally limited by the model of
Natural Language Processing implemented in NetBase®.
3.3.1. KH Coder
This is a software that analyzes text based content quantitatively in service of computational
linguistics (Higuchi, 2014). Furthermore it can be utilized for the generation of relationships
between words in text which can then be clustered among close or like terms or features (Naoi, et
al., 2011). KH coder can be used to perform the following analysis.
1. Compute co-occurrence network of words. That is bringing out words that have the
tendency to occur several times and co-relate. It also has the capability to select specific
key words to show the co-occurrence of a flow, as shown in Fig. 8.
2. Analyze the correspondence words. This is a statistical technique used to analyze multi-
way or two-way tables that contain a measure of correspondence between the rows and
columns. It is an exploratory data analysis tool that is used to identify systematic relations
between variables in the absence of a priori hypothesis about the relationship between such
variables. It resembles traditional hypothesis testing. KH Coder also provides options to
filter the display of words by Chi-square value, grey-scaled dot, bubble plot (unesco.org,
2008)
38
Fig. 7: Illustrating the association of words with other words used with them, an example of co-occurrence network of
words. Source: (Higuchi, 2013)
Fig. 8: Illustrating the association of words with correspondence analysis. Source: (Higuchi, 2013)
39
3. Hierarchical cluster analysis uses a variety of levels to group data into a form of cluster
tree. The tree is a multilevel hierarchy in which clusters at one level combine to form
clusters at the next level. Thus it allows the flexibility to control the number of levels most
appropriate for an application.
Fig. 9: An example of output from hierarchical analysis using KH Coder. Source: (Higuchi, 2013)
3.3.2. Raw dataset classification criteria
The results or sentiments are classified into four basic categories:
40
Positive: Positive implies that the comment provided by the individual about a specific
discussion is ’good’. The positive sentiment is assigned to a comment if that particular
comment is factual and the individual agrees with the discussion. For example, if the discussion
topic is ‘Would you purchase an IPhone 5S Phone’? and the individual’s comment to this
question is ‘Yes, I would definitely buy an IPhone 5S’, their response/comment is known as a
positive comment.
Negative: Negative is the opposite of the positive; it implies that the sentiment behind a
comment given in a specific discussion is ‘bad’ and the individual does not agree with the
discussion topic. For example, if the discussion topic is “Would you purchase an IPhone 5S
Phone?” and the individual’s answer/comment to this question is ‘No, I would not definitely
buy an IPhone 5S’. Their response would be a negative comment and the sentiment behind the
comment is negative.
Mixed: This category involves the combination of both positive and negative. In this case the
sentiment behind a comment is neither completely positive nor negative, instead there is some
form of mixed feelings and condition. For example if the discussion topic is ‘Would you
purchase an IPhone 5S Phone?’ and the individual’s comment to this question is ‘I might buy
an IPhone 5S if I have enough money’, that type of comment would be considered as a mixed
comment and the sentiment behind the comment is mixed.
Neutral: In this case the sentiment behind a comment is neither positive nor negative instead
there is no specify answer to the question. For example if a discussion topic is ‘Would you
purchase an IPhone 5S Phone’? If the individuals comment to this question is ‘I am not sure if
I will buy an IPhone 5S’. This comment is known as a neutral comment and the sentiment
behind the comment is neutral.
In summary, the correct application of NetBase® to extract data from social media networks affects
how reliable the analysis of the data will be, where such analysis aims to establish whether the data
is good enough for the determination of important information like the economic confidence of
consumers, quantified as the CCI. In Chapter 4, the data extracted from NetBase® using the
[search] keywords is presented in graphical, qualitative and quantitative forms and analyzed
accordingly. Then fundamental attributes of the data tested in the KH Coder tool are presented.
_________________________
3 “The # symbol, called a hashtag, is used to mark keywords or topics in a Tweet. It was created organically by Twitter
users as a way to categorize messages” (Twitter, Inc., 2014)
41
Chapter 4
Results and Analysis
Each keyword that was used as a search term for the period of 3 months, 6 months and 1 year, resulted in
three different outputs from NetBase. The first is a graphic summary of the extracted dataset NetBase
classified as being associated with the search term. This graphic summary is a word cloud that summarize
highlights of the dataset output for a keyword. For instance, the search term ‘Capital Investment’ for the
period of 6 months, gave an associated word cloud that summarized the ‘Top Emotions’ expressed, the
‘Top Terms’, the ‘Most used hash tags3’, and the ‘Top conversation places’. Please see Fig. 10.
(a) (b)
(c)
Fig. 10: Word cloud of (a) Top Emotions expressed; (b) Top Terms used and (c) Most used hash tags
42
The other output from NetBase is a tabulated summary of quantified measurements of the various insights
given in the word clouds. Table 2 shows the ‘Top Terms’ used alongside the number of times they occurred
in social media networks discussions over the period of interest. The top words used to express emotions
associated with Capital Investments are also shown in Table 2.
Table 2: A frequency count summarizing the Top terms & Top emotion words related to Capital Expenditure
The third output is the raw dataset NetBase® collected from a variety of social media networks and indexed.
There are 500 of these for each of the time periods per searched keyword. It contains the actual Facebook
comments, Twitter tweets, Blog comments and comments from a variable of other sources. These are called
Terms Mentions
investors 1207563
Fund 990841
projects 698835
property 681345
stocks 663855
income 636100
industry 588573
Security 538384
Children 462341
return 438008
U.S. 361204
economy 358923
report 348657
assets 324703
money 320901
sector 313297
capital 312537
dollars 296948
firm 281739
opportunities 279078
costs 270713
loans 268051
interest 267291
rate 260067
value 258166
risk 255124
growth 242957
family 235353
bank 228889
laws 225847
member 192769
portfolio 184404
Profit 179081
education 177560
benefits 172617
cash 168055
interest rates 167294
foreign investment 165393
Federal Government 163492
Investment Bank 162351
good investment 148283
data 140299
policies 133835
Income Tax 126991
equity 122809
bonds 117866
China 116726
Return on Investment 112923
hedge fund 108361
development 103798
Emotions Positive Mentions Negative Mentions
best 57465 0
good 32875 0
great 26819 0
enjoy 9213 0
prefer 7880 0
interested 6658 0
important 6271 0
like 4598 0
successful 4495 0
happy 3831 0
thank 3819 0
love 3796 0
worth it 3615 0
attractive 3539 0
solid 3491 0
not fail 3314 0
worthwhile 2653 0
look forward to 2361 0
appropriate 2235 0
pleased 2102 0
favorable 1896 0
vital 1634 0
beneficial 1632 0
not force 1591 0
proud 1326 0
blame 0 7778
bad 0 4725
concerned 0 3553
worried 0 3028
accuse 0 2722
hate 0 2515
not want 0 1980
not like 0 1790
fear 0 1782
complain about 0 1400
poor 0 1368
worthless 0 1135
not afford 0 1079
worst 0 1029
worse 0 996
wrong 0 993
fuck 0 986
stuck 0 925
bad bitch 0 887
stupid 0 864
shit 0 862
frustrating 0 822
terrible 0 749
fed up 0 745
tired 0 732
43
Sound Bite Text. However, NetBase® also classed these Sound Bite Texts as either being Positive, Negative
or Neutral. Other information include the domain name of the social media network where each Sound Bite
Text is extracted from, date and time stamps of publication among other details, as shown in Table 3.
For each of the keywords used to extract data from NetBase®, these three sets of outputs are obtained for
the 3-month, 6-month and 1-year periods. Given that there are fourteen search terms and nine sets of data
outputs per search term, there are total of 126 datasets to analyze qualitatively and quantitatively. Therefore
it is important to focus on the objective of the thesis—i.e. to evaluate whether data from social media
networks could be used to meet the needs of determining information such as the consumer confidence
index. In light of these constraints, seven of the datasets captured for each searched keyword were analyzed
instead of all fourteen.
The word cloud summary results shown in Fig. 10 were analyzed qualitatively. Then the tabular results
shown in Table 2 were graphically analyzed. The results shown in Table 3 from all fourteen searched
keywords were combined and analyzed collectively using statistical methods implemented in the KH Coder
tool. These gave different rich perspectives on the attributes of the data from social media networks, their
strengths and weaknesses if they were to be used to measure a metric like CCI.
44
Sequ
ence
Soun
d Bite
Text
Soun
d Bite
Sent
imen
tLikes
Disli
kes
Pos E
mot
ions
Neg E
mot
ionsPo
s Beh
avio
rsNeg B
ehav
iors
Title
Sour
ce Ty
peURL
Dom
ainPu
blish
ed D
ate (
GMT-
04:00
) New
York
Auth
or G
ende
rAu
thor
URLA
utho
r Nam
eAuth
or H
andl
eAu
thor
Loca
tion
No. o
f Fol
lowe
rs/Da
ily U
niqu
e Visi
tors
Klou
t Sco
reSim
ilar
1Fo
reign
inve
stmen
t in S
pain
is no
w m
ore a
ttrac
tive.
In Se
ptem
ber,
the S
pani
sh G
over
nmen
t pas
sed a
new
law w
hich
gran
ts au
tom
atic
resid
ency
to fo
reign
inve
stors.
Non
-EU
hous
e buy
ers c
an no
w ga
in a
Span
ish vi
sa fo
r the
mse
lves,
thei
r par
tner
, chi
ldre
n und
er th
e age
of 18
and d
isabl
ed ch
ildre
n of a
ny ag
e.
Posit
iveSo
cial N
etwo
rks
http
://ww
w.fa
cebo
ok.co
m/1
8388
2498
3875
53/p
osts/
4375
0161
6358
972
face
book
.com2013
-10-
19 03
:43:15
http
://ww
w.fa
cebo
ok.co
m/p
rofil
e.ph
p?id
=183
8824
9838
7553
Ibiza
Soth
eby's
Inte
rnat
iona
l Rea
ltyNo
2Th
is is
in sp
ite of
the f
act t
hat S
wedi
sh da
ta tr
affic
is m
uch g
reat
er th
an in
Nor
way,
the w
ebsit
e said
, citi
ng st
atist
ics fr
om N
orwe
gian r
egul
ator
PT. T
he w
ebsit
e put
diffe
renc
e with
Den
mar
k dow
n to t
he to
ugh c
ompe
titio
n the
re. C
ompu
terW
orld
says
Nor
wegia
n ope
rato
rs bl
ame t
he hi
gher
price
s on i
nves
tmen
t req
uire
men
ts. In
the m
ost r
ecen
t qua
rter,
NetC
om's
ARPU
was
SKK 2
51 or
NOK
241,
com
pare
d with
SEK 1
85 fo
r Tel
iaSon
era i
n Swe
den a
nd w
ith SE
K 153
in D
enm
ark.
Tele
nor's
seco
nd qu
arte
r ARP
U in
Nor
way w
as N
OK 28
0, co
mpa
red w
ith an
aver
age o
f NOK
299 i
n 201
2, sa
id Co
mpu
terW
orld
. Ne
gativ
ehi
gher
price
blam
eNo
rweg
ian m
obile
ARP
U hi
gher
than
in D
enm
ark,
Swed
enBl
ogs
http
://ww
w.te
leco
mpa
per.c
om/n
ews/
norw
egian
-mob
ile-a
rpu-
high
er-th
an-in
-den
mar
k-sw
eden
--973
937
tele
com
pape
r.com
2013
-10-
18 07
:4450
No
3In
tere
sting
: Tor
y Cou
ncil b
lames
Tory
Gov
ernm
ent f
or ho
mel
essn
ess.
Wes
tmin
ster c
ounc
il say
s wel
fare
refo
rms a
re ca
usin
g the
hom
eles
snes
s sur
ge - F
ined
£100
,000 f
or ill
egall
y kee
ping
fam
ilies
in Bn
B acc
omm
odat
ion,
Wes
tmin
ster c
ounc
il blam
es th
e gov
ernm
ent's
hous
ing b
enef
it cu
ts. "W
estm
inste
r exp
laine
d to t
he om
buds
man
that
hous
ing b
enef
it ca
ps af
fecte
d 5,20
0 of 5
5,000
priva
te te
nant
s in t
he bo
roug
h, al
mos
t 700
of th
em lo
sing b
etwe
en £2
00 an
d £75
0 a w
eek.
Nega
tive
hom
eles
snes
sbl
ame
Socia
l Net
work
sht
tp://
www.
face
book
.com
/178
1286
1220
0341
/pos
ts/68
2640
6217
4913
5fa
cebo
ok.co
m2013
-10-
18 04
:07:45
http
://ww
w.fa
cebo
ok.co
m/p
rofil
e.ph
p?id
=178
1286
1220
0341
Grow
th - N
ot Cu
tsNo
4Ha
ve th
e SNP
not m
ade g
ood a
lmos
t eve
ryth
ing t
hey h
ave p
rom
ised s
ince
first
gain
ing o
ffice
as a
min
ority
party
> I co
uld m
ake a
very
long
list o
f mad
e goo
d pro
mise
s but
will
conf
ine i
t to j
ust a
few.
Rem
oval
of al
l brid
ge to
lls; D
elive
ry of
the n
ew Fo
rth Cr
ossin
g; Ex
ceed
ed th
e pro
mise
d num
ber o
f Bob
bies
on th
e bea
t; Re
ducti
on of
Scot
tish C
rime f
igure
s to r
ecor
d low
s; Co
uncil
Tax F
reez
e; N
o SNP
Coun
cil Be
droo
m Ta
x evic
tions
; Sco
ttish
econ
omy o
utpe
rform
ing t
he U
K. S
cotti
sh in
ward
inve
stmen
t out
perfo
rmin
g the
UK.
Crac
kdow
n on s
ecta
riani
sm; C
rack
down
on A
lcohi
lism
; Cra
ckdo
wn on
Dom
estic
Vio
lance
; Rec
ord n
umbe
rs of
mod
ern a
ppre
ntice
ship
; Sco
ttish
expo
rts of
such
as W
hisk
y at r
ecor
d lev
els;
Scot
tish P
olice
Force
unde
r one
head
; Sco
ttish
NHS
not p
rivat
ised;
Free
High
er ed
ucat
ion;
Free
Pres
cript
ions
; Fre
e hom
e car
e for
the e
lder
ly; Et
c.etc.
Wha
t is i
t you
don't
trus
t the
m fo
r?Po
sitive
outp
erfo
rm U
KCo
mm
entsh
ttp://
www.
hera
ldsc
otlan
d.co
m/p
oliti
cs/re
fere
ndum
-new
s/ye
s-sco
tland
-to-re
veal-
lates
t-wea
pon-
in-re
fere
ndum
-bat
tle.22
4488
27#c
omm
ent-1
0855
6996
0he
rald
scot
land.
com
2013
-10-
17 07
:07:36
p4ke
dNo
5I d
id he
ar (n
ot su
re w
here
) tha
t you
coul
d get
4 fa
mily
mem
bers
in if
your
resid
ing i
n Can
ada i
f you
mak
e the
m co
me o
ver w
ith a
holid
ay vi
sa an
d the
n the
y app
ly fo
r a re
siden
tial v
isa 5
year
s lat
er or
som
ethi
ng? M
y fam
ily (a
nd I d
efin
itely
will b
y the
n to s
uppo
rt m
y sel
f) ha
ve su
fficie
nt fu
nds t
o buy
a ho
use W
ITH ch
ange
. (Th
e hou
se pr
ices i
n Nov
a Sco
tia ar
e rid
iculo
usly
chea
p com
pare
d to t
he U
K) A
nywa
y, th
anks
for t
he ad
vice
Posit
iveRe
: Mov
ing t
o Can
ada w
ith fa
mily
of 4?
(Inc
ludi
ng m
e)Fo
rum
sht
tp://
briti
shex
pats.
com
/foru
m/s
howt
hrea
d.ph
p?t=
8122
46#p
ost1
0946
542
briti
shex
pats.
com
2013
-10-
16 05
:28ht
tp://
briti
shex
pats.
com
/foru
m/m
embe
r.php
?u=2
1723
0As
pirin
gCan
adian
No
6Bo
ttom
line,
my i
nves
tmen
ts gr
ew m
uch g
reat
er un
der O
bam
a tha
n Bus
h, Jr
.Po
sitive
grea
ter u
nder
Oba
ma
Com
men
tshttp
://ww
w.na
tiona
lrevie
w.co
m/c
orne
r/361
326/
hous
e-go
p-m
ay-ca
ncel
-vot
e-ro
bert-
costa
#com
men
t-108
3749
223
natio
nalre
view.
com
2013
-10-
15 19
:02:51
agwt
ubNo
7I t
hink
that
mak
es it
quite
a re
laxed
hous
ing m
arke
t." M
eanw
hile
, in Sw
itzer
land t
here
was
even
mor
e com
petit
ion f
or th
e tax
dolla
r give
n peo
ple p
ay th
ree i
ncom
e tax
es - a
fede
ral o
ne, o
ne to
cant
ons a
nd on
e to c
ounc
ils - y
et st
ill ha
ve in
com
e tax
rate
s am
ong t
he lo
west
in th
e OEC
D. S
witze
rland
has a
lso ha
d sta
ble h
ouse
price
s. Ac
cord
ing t
o Har
twich
Switz
erlan
d offe
rs a "
drea
m sc
enar
io" w
ith pl
anne
rs an
d pol
iticia
ns fa
cing t
he fu
ll cos
ts an
d ben
efits
of th
eir d
ecisi
ons.
For N
Z 'th
e wor
st m
ay be
yet t
o com
e' .
Posit
iveCh
angin
g NZ's
loca
l gov
t fun
ding
mod
el to
repl
icate
Ger
man
y or S
witze
rland
's wo
uld h
elp s
olve
our h
ousin
g woe
s, sa
ys N
Z Ini
tiativ
e exe
cutiv
e dire
ctor
Blog
sht
tp://
www.
inte
rest.
co.n
z/pr
oper
ty/6
6860
/cha
ngin
g-nz
s-loc
al-go
vt-fu
ndin
g-m
odel
-repl
icate
-ger
man
y-or
-switz
erlan
ds-w
ould
-hel
p-so
inte
rest.
co.n
z20
13-1
0-15
00:22
:18Ga
reth
Vau
ghan
New
Zeala
ndNo
8M
ore t
han 1
,200 p
eopl
e are
arre
sted i
n Rus
sian r
aids t
arge
ting m
igran
ts, fo
llowi
ng ri
ots i
n Mos
cow
spar
ked b
y the
mur
der o
f an e
thni
c Rus
sian.
1 vi
ews C
ateg
ory:
Gene
ral >
Euro
pe .
UK's
Osb
orne
woo
s Chi
na w
ith sh
ared
inve
stmen
t . (W
orld
News
Euro
pe) .
Ger
man
'luxu
ry bi
shop
' in Po
pe pl
ea .
Po
sitive
Migr
ants
arre
sted i
n Mos
cow
raid
sBl
ogs
http
://ww
w.m
yhea
dlin
ez.co
m/1
2890
974/
Migr
ants-
arre
sted-
in-M
osco
w-ra
ids/
myh
eadl
inez
.com
2013
-10-
14 06
:40:21
BBC N
ews E
urop
eNo
9&&
det l
il inv
estm
ent w
as a
smar
t cho
ice ..
Ahh
I'm fu
ckin
w/ i
t 😁.
Posit
ivesm
art c
hoice
Micr
oblo
gshttp
://tw
itter
.com
/_Ta
pOut
_City
y/sta
tuse
s/38
8695
3760
5355
1104
twitt
er.co
m2013
-10-
11 11
:59:33
http
://tw
itter
.com
/_Ta
pOut
_City
y†_
Prin
ceBO
SS_Tap
Out_
City
yTe
nnes
see,
Uni
ted S
tate
s11
1835
No
10Ho
use p
rices
in Lo
ndon
are a
bsol
utel
y rid
iculo
us! B
uyin
g som
ewhe
re on
your
own i
s nea
rly im
poss
ible
unle
ss yo
u ear
n buc
ket-l
oads
of m
oney
ever
y mon
th. I
gave
up in
the e
nd an
d mov
ed to
the C
arib
bean
whe
re m
y qua
lity o
f life
impr
oved
by ab
out 4
00%!
Ne
gativ
eCo
mm
entsh
ttp://
www.
frenc
hyum
mym
umm
y.com
/201
3/10
/flip
ping
-hou
ses-c
owbo
ys-a
re-b
ack.h
tml#c
omm
ent-1
0787
1096
5fre
nchy
umm
ymum
my.c
om20
13-1
0-11
08:52
:09ad
0n91
No
11Ho
use p
rices
in Lo
ndon
are a
bsol
utel
y rid
iculo
us! B
uyin
g som
ewhe
re on
your
own i
s nea
rly im
poss
ible
unle
ss yo
u ear
n buc
ket-l
oads
of m
oney
ever
y mon
th. I
gave
up in
the e
nd an
d mov
ed to
the C
arib
bean
whe
re m
y qua
lity o
f life
impr
oved
by ab
out 4
00%!
Ne
gativ
eFli
ppin
g Hou
ses:
The C
owbo
ys A
re Ba
ck!
Com
men
tshttp
://ww
w.fre
nchy
umm
ymum
my.c
om/2
013/
10/fl
ippi
ng-h
ouse
s-cow
boys
-are
-bac
k.htm
l?sho
wCom
men
t=13
8149
5930
481#
c702
2919
2316
7094
9388
frenc
hyum
mym
umm
y.com
2013
-10-
11 08
:52:09
http
s://p
lus.g
oogle
.com
/109
1451
9035
8482
9756
60An
n-M
arie
Sam
pson
Yes
12No
w th
ey're
tryin
g to d
ilute
, del
ay, c
onfu
se an
d do a
nyth
ing o
ther
than
have
a pr
oper
disc
ussio
n abo
ut w
hat's
actu
ally n
eede
d. T
he he
ad of
the E
U ex
ecut
ive ha
s acc
used
the b
loc's
mem
ber s
tate
s of d
oing
too l
ittle
to ke
ep pa
ce w
ith th
eir o
verse
as ri
vals
in di
gital
serv
ices a
nd te
leco
mm
unica
tions
. Jos
e Bar
roso
said
EU in
vestm
ent w
as in
suffi
cient
. (30
.09.20
13) .
Eur
opea
n Uni
on ci
tizen
s hav
e no p
rote
ction
from
NSA
activ
ities
wha
tsoev
er, n
et ac
tivist
Jere
mie
Zim
mer
man
n tol
d DW
- and
he is
back
ed up
by a
new
repo
rt pr
esen
ted t
o the
EU pa
rliam
ent.
Nega
tive
'Snow
den's
reve
latio
ns ha
ve ha
d ver
y litt
le im
pact'
Blog
sht
tp://
www.
dw.d
e/sn
owde
n-s-r
evel
atio
ns-h
ave-
had-
very
-littl
e-im
pact/
a-17
1433
55?m
aca=
en-rs
s-en-
all-1
573-
rdf
dw.d
e20
13-1
0-07
19:43
Germ
any
6988
No
13Ha
san M
alek,
Dom
estic
Trad
e and
Cons
umer
Affa
irs m
inist
er, h
intin
g hat
Putra
jaya m
ay sl
ash s
ubsid
ies f
or fl
our a
nd su
gar i
n Bud
get 2
014 .
Inv
estm
ent f
rom
Chin
a . W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia. D
r Mah
athi
r Moh
amad
, on h
is m
eetin
g with
Chin
a's Pr
esid
ent X
i Jin
ping
.
Posit
iveHA
RAKA
H Ho
rses'
Mou
ths
Blog
sht
tp://
hafif
i-haf
iz.bl
ogsp
ot.co
m/2
013/
10/h
arak
ah-h
orse
s-mou
ths_
7.htm
lha
fifi-h
afiz.
blog
spot
.com
2013
-10-
07 06
:15ht
tp://
www.
blog
ger.c
om/p
rofil
e/05
6735
1962
2677
6103
3Ad
min
No
14Ha
san M
alek,
Dom
estic
Trad
e and
Cons
umer
Affa
irs m
inist
er, h
intin
g hat
Putra
jaya m
ay sl
ash s
ubsid
ies f
or fl
our a
nd su
gar i
n Bud
get 2
014
. In
vestm
ent f
rom
Chin
a .
We l
ike to
see m
ore C
hine
se in
vestm
ents
in M
alays
ia. A
t the
mom
ent,
Mala
ysia'
s inv
estm
ent i
n Chi
na is
mor
e tha
n Chi
nese
inve
stmen
ts in
Mala
ysia.
Dr M
ahat
hir M
oham
ad, o
n his
mee
ting w
ith Ch
ina's
Pres
iden
t Xi J
inpi
ng .
Po
sitive
BERI
TA TE
RBAR
U -->
Unc
ivilis
ed A
ttitu
de - H
orse
s’ M
outh
s – 7
Octo
ber 2
013:
Suga
...So
cial N
etwo
rks
http
://ww
w.fa
cebo
ok.co
m/h
arak
ahda
ily/p
osts/
1015
1791
3517
4836
9fa
cebo
ok.co
m2013
-10-
06 23
:42:21
Hara
kahd
aily
Yes
15"S
o, he
(Xi)
will c
onsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d. M
ore i
nves
tmen
t sou
ght O
ther
issu
es di
scus
sed i
n the
hour
-long
mee
ting i
nclu
ded i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Mah
athi
r. "W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia," h
e add
ed. M
ahat
hir h
ad al
so in
vited
Xi to
enco
urag
e mor
e Chi
nese
to in
vest
in th
e Rub
ber C
ity an
d the
Aut
omot
ive Co
mpo
nent
s City
, to b
e bui
lt at
Kota
Putra
, situ
ated
at th
e Mala
ysia-
Thai
bord
er.h
ttp://
www.
mala
ysiak
ini.c
om/n
ews/
2430
80 Tu
n Dr M
ahat
hir M
oham
ad to
day s
aid M
alays
ians s
houl
d not
forg
et th
at th
e lat
e com
mun
ist le
ader
Chin
Peng
had w
ante
d to s
et up
a co
mm
unist
stat
e in t
he co
untry
. Po
sitive
見高拜
,見低踩
,mam
ak 簡直是勢利小人
Foru
ms
http
://cfo
rum
.cari.
com
.my/
foru
m.p
hp?m
od=v
iewt
hrea
d&tid
=321
8889
&ext
ra=p
age%
3D1#
pid1
1382
3116
cforu
m.ca
ri.co
m.m
y20
13-1
0-06
19:32
(http
://cfo
rum
.cari.
com
.my/
hom
e.ph
p?m
od=s
pace
cp&a
c=pr
ofile
&op=
verif
y&vid
=1,h
ttp://
cforu
m.ca
ri.co
m.m
y/ho
me.
php?
mod
=spa
ce&u
id=2
0340
8)jvc
pcv5
5M
alays
iaNo
16"S
o, he
(Xi)
will c
onsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d. O
ther
issu
es di
scus
sed i
n the
hour
-long
mee
ting i
nclu
ded i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Dr M
ahat
hir.
"W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia," h
e add
ed. D
r Mah
athi
r had
also
invit
ed Xi
to en
cour
age m
ore C
hine
se to
inve
st in
the R
ubbe
r City
and t
he A
utom
otive
Com
pone
nts C
ity, t
o be b
uilt
at Ko
ta Pu
tra, s
ituat
ed at
the M
alays
ia-Th
ai bo
rder
. Po
sitive
Whe
n a sm
art p
ower
ful C
hine
se m
et a
smar
t Mel
ayu o
ldm
anBl
ogs
http
://lif
eofa
anni
e.bl
ogsp
ot.co
m/2
013/
10/w
hen-
smar
t-pow
erfu
l-chi
nese
-met
-smar
t.htm
llif
eofa
anni
e.bl
ogsp
ot.co
m20
13-1
0-05
21:07
Fem
aleht
tp://
www.
blog
ger.c
om/p
rofil
e/17
6655
3603
3017
7676
90An
nie
Yes
17"S
o, he
(Xi)
will c
onsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d. O
ther
issu
es di
scus
sed i
n the
hour
-long
mee
ting i
nclu
ded i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Dr M
ahat
hir.
"W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia," h
e add
ed.
Posit
iveBl
ogs
http
://ww
w.th
ebor
neop
ost.c
om/2
013/
10/0
6/dr
-mah
athi
r-pro
pose
s-ass
ociat
ion-
in-m
emor
y-of
-adm
iral-c
heng
-ho/
theb
orne
opos
t.com
2013
-10-
05 13
:01:48
341
Yes
18"S
o, he
(Xi)
will c
onsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d. O
ther
issu
es di
scus
sed i
n the
hour
-long
mee
ting i
nclu
ded i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Dr M
ahat
hir.
"We l
ike to
see m
ore C
hine
se in
vestm
ents
in M
alays
ia. A
t the
mom
ent,
Mala
ysia'
s inv
estm
ent i
n Chi
na is
mor
e tha
n Chi
nese
inve
stmen
ts in
Mala
ysia,
" he a
dded
. Dr M
ahat
hir h
ad al
so in
vited
Xi to
enco
urag
e mor
e Chi
nese
to in
vest
in th
e Rub
ber C
ity an
d the
Aut
omot
ive Co
mpo
nent
s City
, to b
e bui
lt at
Kota
Putra
, situ
ated
at th
e Mala
ysia-
Thai
bord
er.
Posit
iveDR
MAH
ATHI
R PRO
POSE
S ASS
OCIA
TION
IN M
EMOR
Y OF A
DMIR
AL CH
ENG
HOBl
ogs
http
://kw
ongw
ah.co
m.m
y/ne
ws/2
013/
10/0
5/95
.htm
lkw
ongw
ah.co
m.m
y20
13-1
0-05
08:40
:36M
alays
ia55
Yes
19So
, he (
Xi) w
ill co
nsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d.M
ore i
nves
tmen
t sou
ghtO
ther
issu
es di
scus
sed i
n the
hour
-long
mee
ting i
nclu
ded i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Mah
athi
r."W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia," h
e add
ed.M
ahat
hir h
ad al
so in
vited
Xi to
enco
urag
e mor
e Chi
nese
to in
vest
in th
e Rub
ber C
ity an
d the
Aut
omot
ive Co
mpo
nent
s City
, to b
e bui
lt at
Kota
Putra
, situ
ated
at th
e Mala
ysia-
Thai
bord
er.-
Posit
iveSo
cial N
etwo
rks
http
://ww
w.fa
cebo
ok.co
m/5
1325
4732
0490
03/p
osts/
5955
8822
3815
653
face
book
.com2013
-10-
05 07
:14:15
http
://ww
w.fa
cebo
ok.co
m/p
rofil
e.ph
p?id
=513
2547
3204
9003
Tima I
nter
natio
nal
Yes
20"S
o, he
(Xi)
will c
onsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d. O
ther
topi
cs di
scus
sed i
n the
hour
-long
mee
ting w
ere i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Dr M
ahat
hir.
"W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia," h
e add
ed. C
hina
's de
cisio
n to o
pen c
onsu
l-gen
eral
in Pe
nang
show
s sta
te's
impo
rtanc
e, sa
ys ch
ief m
inist
er Pr
esid
ent X
i call
s Mala
ysia
a tru
stwor
thy s
trate
gic pa
rtner
- Ber
nam
a Cha
lleng
ers n
eed t
o be g
iven a
chan
ce in
party
polls
, say
s Kha
iry - B
erna
ma C
hina
Pres
iden
t end
s visi
t to M
alays
ia, fl
ies o
ff to
Bali f
or A
PEC m
eet -
Bern
ama A
hmad
Zahi
d get
s bigg
est c
heer
s at U
mno
dele
gate
s' se
ssio
ns, n
ot so
Hish
amm
uddi
n For
mer
depu
ty pr
esid
ent c
ritici
ses U
mno
for s
ides
tepp
ing l
eade
rship
cont
est
Posit
iveDr
M su
gges
ts M
alays
ia-Ch
ina t
rade
asso
ciatio
n dur
ing m
eetin
g with
Xi Ji
npin
g - Be
rnam
aBl
ogs
http
://all
news
.rss.t
hem
alays
ianin
sider
.com
/c/3
3362
/f/56
7634
/s/3
2154
8aa/
sc/3
9/l/0
L0St
hem
alays
ianin
sider
0N0C
busin
ess0
Carti
cle0C
dr0E
m0E
sugg
ests0
Emala
ysia0
Echi
na0E
trade
0Eas
socia
tion0
Edur
ing0
Emee
ting0
Ewith
0Exi0
Ejinp
i/sto
ry01
.htm
allne
ws.rs
s.the
mala
ysian
insid
er.co
m20
13-1
0-05
06:20
:48ed
itor@
them
alays
ianin
sider
.com
587
Yes
21"S
o, he
(Xi)
will c
onsid
er th
is be
caus
e he t
hink
s thi
s is a
good
idea
and h
e will
see w
hat h
e can
do to
supp
ort i
t," he
adde
d. O
ther
topi
cs di
scus
sed i
n the
hour
-long
mee
ting w
ere i
nten
sifyin
g tra
de vo
lum
e bet
ween
both
coun
tries
, said
Dr M
ahat
hir.
"W
e like
to se
e mor
e Chi
nese
inve
stmen
ts in
Mala
ysia.
At t
he m
omen
t, M
alays
ia's i
nves
tmen
t in C
hina
is m
ore t
han C
hine
se in
vestm
ents
in M
alays
ia," h
e add
ed. C
hina
's de
cisio
n to o
pen c
onsu
l-gen
eral
in Pe
nang
show
s sta
te's
impo
rtanc
e, sa
ys ch
ief m
inist
er Pr
esid
ent X
i call
s Mala
ysia
a tru
stwor
thy s
trate
gic pa
rtner
- Ber
nam
a Chi
na Pr
esid
ent e
nds v
isit t
o Mala
ysia,
flie
s off
to Ba
li for
APE
C mee
t - Be
rnam
a Ahm
ad Za
hid g
ets b
igges
t che
ers a
t Um
no de
lega
tes'
sess
ions
, not
so H
isham
mud
din F
orm
er de
puty
pres
iden
t crit
icise
s Um
no fo
r sid
este
ppin
g lea
dersh
ip co
ntes
t Chi
na's
decis
ion t
o ope
n con
sul-g
ener
al in
Po
sitive
Dr M
sugg
ests
Mala
ysia-
Chin
a tra
de as
socia
tion d
urin
g mee
ting w
ith Xi
Jinp
ing
Blog
sht
tp://
allne
ws.rs
s.the
mala
ysian
insid
er.co
m/c
/333
62/f/
5676
34/s
/321
5683
0/sc
/39/
l/0L0
Sthe
mala
ysian
insid
er0N
0Cbu
sines
s0Ca
rticle
0Cdr
0Em
0Esu
gges
ts0Em
alays
ia0Ec
hina
0Etra
de0E
asso
ciatio
n0Ed
urin
g0Em
eetin
g0Ew
ith0E
xi0Eji
npi/s
tory
01.h
tmall
news
.rss.t
hem
alays
ianin
sider
.com
2013
-10-
05 06
:20:48
edito
r@th
emala
ysian
insid
er.co
m58
7Ye
s
22Th
is is
the p
robl
em th
at go
vern
men
t has
crea
ted.
We d
on't h
ave f
amili
es an
ymor
e, th
e gov
ernm
ent h
as ta
ken t
hat p
ositi
on ov
er. T
his i
s wha
t hap
pens
whe
n we k
eep h
andi
ng ou
t to t
he po
or, t
hey s
tay p
oor.
Nega
tive
creat
e pro
blem
Com
men
tshttp
://m
oney
.cnn.
com
/201
3/10
/04/
news
/eco
nom
y/sh
utdo
wn-h
ead-
start/
#com
men
t-107
0200
909
mon
ey.cn
n.co
m20
13-1
0-04
10:13
:0602
w84
1507
47No
23RT
@_G
_SW
AGG:
I'm no
t per
fect,
but I
'm w
orth
it. G
reat
est i
nves
tmen
t you
'll ev
er m
ake.
Posit
iveM
icrob
logsh
ttp://
twitt
er.co
m/a
mor
ttay/
statu
ses/
3858
9290
4129
9374
08tw
itter
.com20
13-1
0-03
18:23
:30ht
tp://
twitt
er.co
m/a
mor
ttay
E.K.
amor
ttay
917
40No
24le
ss th
an ha
lf th
e pric
e of a
sing
le Sp
ace S
huttl
e flig
ht? O
n sec
ond t
houg
ht if
you c
alled
it an
inve
stmen
t. I t
hink
its t
he be
st in
vestm
ent t
he U
SA ev
er m
ade.
but a
ctuall
y an i
nves
tmen
t im
ply s
ther
e is a
risk
of m
oney
... a
FIXED
PRIC
E Mile
stone
paym
ent i
s no r
isk.
Posit
iveCo
mm
entsh
ttp://
www.
mot
herjo
nes.c
om/n
ode/
2306
71#c
omm
ent-1
0693
6921
4m
othe
rjone
s.com
2013
-10-
03 17
:39:30
zhsc
i18
657
No
25Bu
sines
s inv
estm
ents
in A
ustra
lia is
grea
t.Po
sitive
Com
men
tshttp
://ww
w.th
ebor
neop
ost.c
om/2
013/
09/3
0/pr
oper
ty-m
arke
t-upd
ate-
iden
tifie
s-goo
d-tim
e-to
-inve
st-in
-mel
bour
ne-p
erth
-pro
perti
es/#
com
men
t-106
8439
371
theb
orne
opos
t.com
2013
-10-
03 04
:38:51
ad04
w234
1No
26W
e hav
e peo
ple f
eelin
g goo
d abo
ut no
t was
ting e
nerg
y but
man
ufac
turin
g is t
he gr
eate
st en
ergy
cons
umer
, and
we n
eed m
ore m
anuf
actu
ring.
Econ
omizi
ng is
a sp
iritu
al ac
t, no
t a m
ater
ial on
e.) C
hina
is m
akin
g pro
ducti
ve in
vestm
ents.
We a
re m
akin
g red
istrib
utive
expe
nses
, and
spen
ding
trill
ions
defe
ndin
g oil a
nd tr
ade r
oute
s, an
d our
prim
ary e
xpor
t - th
e dol
lar. A
nd w
e will
not g
et an
ywhe
re th
inkin
g tha
t som
e ver
y sm
all m
inor
ity of
a Co
nfuc
ian po
pulat
ion,
or ou
r odd
obse
ssio
n with
the r
eligi
on of
Uni
versa
l Dem
ocra
tic Se
cular
Hum
anism
will
have
any l
ong t
erm
effe
ct on
the S
inic
cultu
re.
Posit
iveSo
cial N
etwo
rks
http
://ww
w.fa
cebo
ok.co
m/7
4119
7263
/pos
ts/10
1519
2281
8462
264
face
book
.com2013
-10-
02 06
:16:36
Male
http
://ww
w.fa
cebo
ok.co
m/p
rofil
e.ph
p?id
=741
1972
63Cu
rt Do
olitt
leNo
27RT
@m
arta
baus
ells:
"Inv
estm
ent b
ankin
g is a
trap
, a ga
me a
nd an
addi
ction
". Th
e 10 b
est q
uote
s fro
m fi
nanc
ial in
sider
s | @
joris
luye
ndijk
ht...
Mixe
dad
dicti
on, t
rapad
dicti
onM
icrob
logsh
ttp://
twitt
er.co
m/re
boot
dude
/sta
tuse
s/38
5072
5711
3097
4209
twitt
er.co
m2013
-10-
01 12
:03:48
http
://tw
itter
.com
/rebo
otdu
deJa
wwad
Faridre
boot
dudeP
akist
an59
649
No
28Re
cipie
nt Fr
ance
prais
ed th
e dev
elop
men
t of t
he pl
ane a
s a hi
gh-p
erfo
rman
ce te
chno
logic
al fe
at th
at ca
me a
t a co
st. T
he he
ad of
the E
U ex
ecut
ive ha
s acc
used
the b
loc's
mem
ber s
tate
s of d
oing
too l
ittle
to ke
ep pa
ce w
ith th
eir o
verse
as ri
vals
in di
gital
serv
ices a
nd te
leco
mm
unica
tions
. Jos
e Bar
roso
said
EU in
vestm
ent w
as in
suffi
cient
. Ne
gativ
eLu
xury
TV m
aker
Loew
e file
s for
inso
lvenc
yBl
ogs
http
://ww
w.dw
.de/
luxu
ry-tv
-mak
er-lo
ewe-
files
-for-i
nsol
venc
y/a-
1712
9473
?mac
a=en
-rss-e
n-all
-157
3-rd
fdw
.de
2013
-10-
01 07
:58Ge
rman
y69
88Ye
s
29EU
top c
ourt
rule
s aga
inst
rail f
irm in
com
pens
atio
n-fo
r-del
ay di
sput
e . T
he he
ad of
the E
U ex
ecut
ive ha
s acc
used
the b
loc's
mem
ber s
tate
s of d
oing
too l
ittle
to ke
ep pa
ce w
ith th
eir o
verse
as ri
vals
in di
gital
serv
ices a
nd te
leco
mm
unica
tions
. Jos
e Bar
roso
said
EU in
vestm
ent w
as in
suffi
cient
. EU
outli
nes a
mbi
tious
plan
to en
d roa
min
g fee
s and
'thro
ttle'
net n
eutra
lity o
ppon
ents
. US
cons
ultin
g gro
up In
terb
rand
has c
rown
ed co
nsum
er el
ectro
nics
gian
t App
le as
the b
est-p
erfo
rmin
g glo
bal b
rand
in it
s 201
3 ran
king.
Nega
tive
Mar
kets
stay c
alm as
US g
over
nmen
t shu
ts do
wnBl
ogs
http
://ww
w.dw
.de/
mar
kets-
stay-
calm
-as-u
s-gov
ernm
ent-s
huts-
down
/a-1
7129
221?
mac
a=en
-rss-e
n-all
-157
3-rd
fdw
.de
2013
-10-
01 05
:41Ge
rman
y69
88Ye
s
30Ba
rroso
sees
EU m
embe
rs fa
lling
behi
nd in
digit
al wo
rld .
The h
ead o
f the
EU ex
ecut
ive ha
s acc
used
the b
loc's
mem
ber s
tate
s of d
oing
too l
ittle
to ke
ep pa
ce w
ith th
eir o
verse
as ri
vals
in di
gital
serv
ices a
nd te
leco
mm
unica
tions
. Jos
e Bar
roso
said
EU in
vestm
ent w
as in
suffi
cient
. App
le de
thro
nes C
oca C
ola i
n bes
t glo
bal b
rand
s ran
king .
Jap
anes
e con
sum
er el
ectro
nics
com
pany
Tosh
iba h
as an
noun
ced i
t will
clos
e dow
n or s
ell s
ome o
f its
over
seas
tele
visio
n plan
ts ov
er th
e nex
t six
mon
ths.
Nega
tive
Germ
an jo
bles
s rat
e dro
ps on
autu
mn e
cono
mic
pick
upBl
ogs
http
://ww
w.dw
.de/
germ
an-jo
bles
s-rat
e-dr
ops-o
n-au
tum
n-ec
onom
ic-pi
ckup
/a-1
7129
178?
mac
a=en
-rss-e
n-all
-157
3-rd
fdw
.de
2013
-10-
01 05
:02Ge
rman
y69
88Ye
s
Tab
le 3
: T
he
thir
d o
utp
ut
fro
m a
Net
Base
sea
rch
; th
e in
dex
ed d
ata
set
extr
act
ed
fro
m s
oci
al
med
ia n
etw
ork
s
for
the
sea
rch
ter
m o
f C
ap
ital
Ex
pen
dit
ure
45
4.1. Assessing the robustness of social media networks data
The analysis is divided into three sections based on the type of output obtained from NetBase. The first
section is a comparative analysis of the infographic word clouds output by NetBase: Part A. The aim is to
examine how these summaries compare to the mainstream economic discussions during each period of
time. The next part of the analysis follows from the first part, and entails the visualization of the data shown
in Table 2. The aim of the visualization is to explore the aggregate summaries made by NetBase: Part B.
This happens to be the numeric values behind the infographic data from NetBase and shown in Fig. 10. It
gives a quantitative perspective of indicators such as—how many times a vocabulary was used across social
media networks over each period of time and whether there are notable changes in vocabulary used. In Part
C, the largest sources of SMN data were identified, each with its own contribution of content for each
keyword discussed. Then in Part D, the raw data indexed by NetBase was analyzed. This analysis was used
to assess different aspects of how NetBase as a tool performed. For instance how it classed different
Soundbites—positive, negative or neutral—and whether these are deemed the correct classification. This
analysis also looked at the interval between SMNs data extractions and the number of Soundbites that were
repeated, even when they are not from the same source.
In the last part of this analysis, statistical analytical methods were used to explore fundamental attributes of
the raw data from NetBase. This was to check that the classification performed by NetBase meets basic
language semantic rules. For example whether two words that are used together, occurred together in a
Soundbite. This was achieved using the KH Coder tool in section 4.2.
4.1.1. Part A: Qualitative analysis based on Infographics
The infographic summary from NetBase is collated for seven of the fourteen keywords. Each one has ten
word cloud summaries for the ‘Most used hashtags’, ‘Top Attributes’, ‘Top Behaviors’, ‘Top Brands with
the category’, ‘Top Emotions’, ‘Top Likes’, ‘Top Terms’, ‘Top Conversation Places’, proportion of
‘Gender’ contributions and ‘Top Dislikes.’
For the purposes of this research each of these is evaluated based on two criteria to streamline the scope of
analysis and keep it focused. These are (a) whether the category gives an insight into the vocabulary that is
directly associated with the context of the search term and (b) whether it has a contextual relevance to the
expression of human sentiment as it relates to their economic situation. This is because the notion of
consumers’ economic confidence is a sentimental one.
___________________________ 4 A compulsory contribution to state revenue, levied by the government on workers’ income and business profits, or
added to the cost of some goods, services and transactions (Oxford Dictionaries, 2015)
46
Hashtags have become a dominant feature of written human expressions in this era of social media
networks. It is used to associate expressions to any subject of interest, popular or unpopular. Therefore it is
one of the categories that would be analyzed because it meets both criteria. The other categories are ‘Top
emotions’ and ‘Top Terms’ which meet the requirements of both criteria.
From the literature review it is known that human conversations rarely reference the formal name of a
subject of interest being discussed. For instance, people express their opinion on a government policy by
making cynical statements that do not contain a word from the policy document. Or better still they use
swear words or slurs to express how they feel.
This is a constraint for this research because terms such as: consumer confidence, political affiliation,
employment condition or social status, are minimally used terms in informal conversations. Therefore, to
mitigate this limitation, two keywords were chosen—one is used frequently in informal conversations and
the other is used frequently in formal conversations. This choice was made to facilitate a richer analysis.
Tax, a frequently used word (in both informal and formal conversations) is one of the keywords of choice.
The other keyword is Capital expenditure, a frequently used word in formal conversations. It was hoped
that analyzing the data generated by these keywords would throw some light on the strengths and
weaknesses of data from social media networks. The other keywords were analyzed collectively because
they fell into one of either category of being formal or informal. Therefore they will share similar attributes
to either one or the other of the two keywords selected for analysis.
4.1.1.1. Tax
From the literal definition of tax4, everyone has the obligation to pay some form of tax. The word/term is
also used in formal and informal conversations and it attracts direct feedback of opinion, especially in the
event of a change of government tax policies. Various forms of taxes include the Value Added Tax which
impacts on the cost of living via the cost of goods & services; corporation tax—on the profits of companies;
sales tax—usually added to the basic price of a commodity; import-export duty tariff—on goods brought
into or taken out of the country or taxation—in the form of the PAYE that is deducted from peoples wages
when they are paid (The New York State Department of Taxation and Finance, 2014) (Investopedia LLC,
2015).
The qualitative outputs for the tax keyword in Fig. 11 identifies words that are related to ‘tax’ and were
used most often. In the category of 'Top Terms' and 'Most used hashtags', these are orange in color, while
in the category of 'Top emotions expressed' these are the green words.
___________________________
5 & 6 “Incidental emotions are the emotions we carry with us to a decision that have nothing to do with the decision.”
“…integral emotions are emotions that are caused by the decision itself.” (Karen Christensen - Forbes India, 2014)
47
Their font size represents their relative frequency of use. A visual inspection reveals that the top terms
associated with the Tax keyword are vocabulary related to our monetary economic system, its institutions
and activities pertinent to their operational existence.
Starting with the key economic player i.e. the Government (the top term), to activities that influence the
economic situation of everyone at a personal level; e.g. commercial or public sector ‘projects’ that generate
cash flow across the economy, ‘investment’ in infrastructure by Government , ‘funding’ of a variety of
programs as part of government policy, ‘income’ and ‘pension plans’. There are also entities that influence
the state of the economy through the use of economic levers. These are the ‘banks’ and their ‘interest rates’,
political parties—‘Republicans’ or ‘Democrats’ and how their leadership of government could mean a
change in ‘government funding’ policy, exchange rate of the ‘dollars’, the ‘costs’ of running government
departments or the plan for ‘debt’ management.
So far as a government is concerned, taxation is a fundamental mechanism to incentivize or deter economic
actions of any kind. This is because the government’s policy programs, departments, projects and labor
must be financed. What this means for other economic players—from corporations to small businesses and
individuals- is that they are obliged to fund the government through tax. Given that tax is one income
channel used by governments to finance their recurrent expenditure, like debts, salaries, pensions and
capital expenditure, tax affects the economic outlook of everyone in an economy.
Agreeably, these ‘Top terms’ associated with Tax as an aspect of a consumer’s economic perspective do
include popular terms used in mainstream economic discussions. The terms are contextually relevant,
however they only give an indication of the popular terms related to Tax, not how consumers feel about
‘tax’, which in turn will affect how they feel about their economic situation.
Even if the “Most used hashtags” were visually inspected for a hint of the general mood at the time, it would
be found lacking. This is due to the fact that the most used hashtags—'#funding', '#savings', or '#cash' give
a non-directional perspective relative to human situations. In that they are identified implies the topics of
most interest in that period, but give no clue as to whether people had more funds available at their disposal
or whether there were more talks of need for funding.
Although data from social media networks easily provides us with insights into what is being said by those
who participate in it, it may not give us the context or intentions of their interactions. This is attributable to
the two forms of emotions which humans experience—the incidental5 and integral6 forms of emotions, and
how they influence our judgment of any situation, including our own (Daas P.J.H and Puts M.J.H -
European Central Bank, 2014). This is where new tools like NetBase® are of crucial help.
48
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 11: Qualitative summary attributes of data associated with Tax keyword
_____________________
On close observation, these summary infographics seem to originate mainly from North America given the type of terms seen.
For example there is the #tcot and #obamacare hashtags; then ‘Republicans’, ‘Democrats’, ‘Bush’ and ‘Senate’ Top terms
synonymous with North America, the U.S.A in particular. Henceforth, it will be assumed that NetBase captured data from the
North American region.
49
Incidental and integral forms of emotions are responsible for the lag in the expression by Dutch users of
Facebook and Twitter SMNs (Daas and Puts 2014). In the same work, they refer to an earlier work where
50% of the data acquired from a major social network platform, Twitter, is deemed "pointless babble." This
prospectively forewarns that some of the data from SMNs is noise, and without contextual relevance to the
subject of interest (Daas P.J.H and Puts M.J.H - European Central Bank, 2014).
These findings are further buttressed by a Facebook research about emotional contagion (Kramer, et al.,
2014). It found that people tend to post positive or negative looking conversations depending on the overall
“feel” of what their social media friends express and the most recent news feed from their Facebook network
on their Home page.
Nevertheless, it is inadmissible to dismiss half of the dataset we have as irrelevant because they are
associated with economic discussions we all participate in, one way or another. Thus, the missing
knowledge here is how to tell the level of accuracy of the model that was used to arrive at this 50% noise
claim.
With this awareness, we now look to know how consumers feel about ‘tax’. We look to the ‘Top emotions
expressed’ for each period as shown in Fig. 11. Not only are the words ‘great’, ‘enjoy’, ‘good’, ‘love’,
‘best’, amazing’ striking as being positive expressions, overall the high proportion of greens (positives)
compared to red (negatives), imply a positive economic outlook by the largest portion of the sample.
One could postulate that the positive expressions could be induced by some ongoing incentives highlighted
in the ‘Top terms.’ For example, financial ‘Benefits’ in form of tax credits given to individuals could induce
the expression of ‘good’ times, a time to ‘enjoy’ favorable ‘investment’ plans by private corporations.
In as much as some causal association could be postulated to claim that the Top emotions expressed tell us
that consumers are feeling happier or more confident— economically based on the Top terms used and the
events surrounding them— the data does not actually support such postulate. Instead, the data gives a linear
association between the most expressed emotions found in social media networks interactions that are
associated with Tax, the searched Keyword.
50
Let us look at the ‘Most used hashtags’ once again. Ignoring observed duplications, hashtags such as
#funding, #cash, #savings and #invest do point at a fact that there was a buzz on social media networks
about the economy and Tax. The implicit requirement to “trust” the classification done by NetBase based
on fundamental Natural Language Processing principles is tempting though attributing relationship is
always dangerous. It could be that the data classified was either contextually relevant to the searched
keyword, or contextually relevant to a higher level of knowledge, in this case the [sentimental] economic
confidence of consumers who are also users of SMNs. An external validation of the ‘mood’ of consumer
sentiments around each period of interest is one way to evaluate these limitations.
This could be achieved by using published CCIs between July and October 2013. However, this research
does not measure consumer confidence, and therefore such external validation of measurements of
Consumer Confidence is excluded from this analysis. Nonetheless, this need for external validation points
at a weakness of lack of causal association beyond the direct association with a keyword used to search.
In summary, unlike CCSs where the context for questions posed to correspondents has a set and clear scope,
interactions on social media networks are diverse, varied and random in nature. Although corporate
institutions and businesses tend to post communications that are directly linked to a subject of interest, this
is not the case for individuals. People tend to post in response to an event or topic of interest being discussed.
They do not necessarily state how such events or topics affect them personally—at that moment or how it
will in the future.
Based on the qualitative analysis of the Tax keyword above, an attribute that cannot be taken for granted
with data from social media networks is that the use of the right terms will yield the extraction of the right
data, where the challenge is for the user to develop a model of association of the search terms with the
knowledge they seek to derive from the data.
From the perspective of the keyword “tax” and its effect on the economic confidence of consumers, two
observations have been identified from the analysis of Fig. 11: (i) that the data from social media networks
is to a degree contextually associated with the searched keyword but (ii) give no indication of causality
between the data and the higher order knowledge for which it is needed. This poses a fundamental question
of how reliable social media networks data could be used beyond the linear association of contextual
vocabulary.
This suggest a review of the results of analysis for the other keyword that is used extensively in formal
discussions—Capital Expenditure.
51
4.1.1.2. Capital Expenditure
This keyword, unlike the Tax keyword, has a longer term footprint and induces a different kind of consumer
reaction. It is described in Investopedia as:
"Funds used by a company to acquire or upgrade physical assets such as property, industrial
buildings or equipment. This type of outlay is made by companies to maintain or increase the
scope of their operations. These expenditures can include everything from repairing a roof to
building a brand new factory" (Investopedia, LLC, 2015).
Capital expenditure spans beyond the precinct of private businesses or organizations. It is also undertaken
by governments, international financial institutions and non-governmental institutions amongst others.
A visual analysis of the data yields these insights. The key insight that when it comes to the Capital
Expenditure, government is the key player (in the ‘Top term’). It can also be seen that factors that influence
capital expenditure investments are also contained in the social media network interactions.
In the 3-month, 6-month and 12-month periods, such factors identified are ‘fuel prices’, ‘oil price’,
‘subsidy’, ‘inflation’, ‘house prices’ or the ‘dollar’—all of which are referenced when governments or
organizations announce strategic, long term investment decisions. For instance the drop in oil prices to as
low as $50/barrel has led major oil & gas companies to even go as far foregoing the development of
lucrative facilities recently. Shell also recently scrapped its $6.5 Billion USD project in Qatar as a result
(Sergie, M.A.; Kayakiran, F.; Bloomberg, 2015). These actions have ramifications for consumers because
cash flow generated by new projects would have reached host communities and create more jobs and new
income streams.
A glance at the Top emotions expressed in Fig. 12 gives the impression that things are ‘cheaper’; that there
is a ‘low gold price’. Also it can be seen that the underlying theme here is the costs of energy being an
influence on capital expenditure decisions. A visual evaluation of the Top Emotions expressed shows
negative sentiment (in red), and is collectively higher than positive ones (green)—heralded by the high
occurrence of the word ‘heating cost’.
Assuming there is contextual association, the ‘top emotions’ expressed hint that consumers are unhappy
about the cost of energy and this is validated by the ‘top terms’ shown in Fig. 12. Furthermore, the Most
Used hashtags point at topics of highest interest, which includes government operations (#shutdown,
#obamacare), energy (#gas, #syria) and elections (#ronpaul, U.S presidential candidate 2012).
52
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 12: Qualitative summary attributes of data associated with the Capital Expenditure keyword
53
Figure 12 could be interpreted to mean that energy prices are unfavorable and consumers are unhappy about
it. On the other hand the majority of the topics of popular interest dwell on subjects other than on the Top
term ‘government’. Based on this lack of clarity, these hidden links of causality or lack of it, are an
inevitable challenge for anyone who intends to use this sort of data to derive knowledge of a higher order.
Without a causal hint to tell whether Top emotions expressed relate to how consumers feel about their
economic situation, social media networks data of an unaligned context could be used. In other words
without knowledge of the context of the interactions from which the data originated, the implicit bias of the
data might be counter useful for the purposes for which it was extracted in the first place.
Irrespective of the bias of the keywords used to extract data from social media networks, the right keywords
will generate any sets of data that are relevant. The onus is on the researcher to develop a systemic
association between the keywords and the knowledge being sought, with an implicit assumption that
contextual relevance is achieved to some degree through the model of data mining implemented.
If contextual relevance is achieved by up to 50%, application and research shows that the data from social
media networks is 'good enough' to support the synthesis of a higher order level of knowledge—albeit from
a qualitative frame of reference. This was demonstrated in the work of Daas and Puts (2014) for the
European Central Bank. They claim they were able to measure the economic confidence of consumers
based on data from Facebook and Twitter, with a correlation of r = 0.9.
In 2013, the Netherlands was identified as having the most users on social media networks in Europe. In
relative terms, about 65% of its population use social media networks (Office of National Statistics, 2013)
(Error! Reference source not found.). With such high proportion of SMN users, if it is assumed that at
least 80% of the 65% actively use social media, then it is reasonable to assume that online conversations
originate from about 50% of the Dutch population. Again, this is relative and not supported by actual data.
In the situation described here, the other half of the population’s economic prospects are unknown to the
researchers. This implies that a categorical statement of accurately measuring consumer confidence based
on social media networks alone cannot be an end by itself. It is rather inconclusive and immaturely
dismissive about the fact that 50% of the population’s views are unknown, and cannot be said to be positive
or negative. The insights gained from analysis of Figures 13 and 14 in subsections 4.1.1.1 and 4.1.1.2, are
deducible from the rest of the data gathered on the other five keywords shown in Fig. 13, Fig. 14, Fig. 15,
Fig. 16 and Fig. 17.
In the same light, the infographic outputs for Employment conditions, Health, Consumer confidence,
Political affiliation & Retirement shown in Figures 13 through 17, show the similar flaws highlighted above.
54
In particular, it can be seen that the ‘Most used hashtags’ do not reflect in the strongest sense, and that the
most used hashtag is related to the searched keyword. For instance, the infographic of Health keyword
shown in Fig. 16 illustrates this. Other infographics however give an assuring alignment with the keyword.
Holistically, the data from social media networks is varied as each searched keyword yields a different
perspective of vocabulary that is different from the others. This is also the case when one looks across each
time period, beyond the most frequent top term, top emotion expressed and most used hashtag.
There is an opportunity to advance models that would provide a second (and higher) order association of
these data with higher levels of informed knowledge.
The analysis undertaken so far has already identified one major challenge in the area of data mining
identified in the literature review: that the accuracy of the data models to associate words to the right context
and to capture the true intentions of the interactions that contain such words on social media networks
represent the key determinants to the usefulness of the data.
In summary, the qualitative analysis of data from social media networks hints that this data is 'good enough'
based on real-time relevance. However this will only hold if the fundamental assumption of contextual
relevance holds.
The next part of the analysis looks into the quantitative aspect of these summarized infographic outputs
from NetBase®. The aim is to assess whether each category such as the ‘top terms’ NetBase® has identified
is supported by the actual number of times such ‘top terms’ occurred. Other attributes of the data that are
looked into includes the volume of online interactions, e.g. tweets, Facebook messages/comments/statuses
and the trend of volume over each period of interest.
55
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 13: Qualitative summary attributes of data associated with the Employment Condition keyword
56
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 14: Qualitative summary attributes of data associated with the Health keyword
57
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 15: Qualitative summary attributes of data associated with the Consumer Confidence keyword
58
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 16: Qualitative summary attributes of data associated with the Political Affiliation keyword
59
3-months
6-months
12-months
Top terms Top emotions expressed Most used hashtags
Fig. 17: Qualitative summary attributes of data associated with the Retirement keyword
60
4.1.2. Part B: Quantitative analysis of aggregated summaries
Tax
The second output from NetBase® for each of the seven keywords selected for analysis is shown in Table
2 (p.42). Analyzing the numeric values yields no hidden trends. As a result, each table obtained for the three
subcategories under review is graphed to visualize the data. It is believed that doing so will enable trends
in the data to be observed easily.
The Top terms associated with the Tax keyword are shown Fig. 18 (a) 3 months, (b) 6 months and (c) 12
months. The reference point here are the first five terms in each period of time and how often they occurred
in the data found on social media networks. Also, the bottom five “top terms” are analyzed to see whether
there any observed movements, such as a change in position and magnitude of occurrence. In a similar
fashion, the top emotions and the most used hashtags are analyzed.
Similar to the qualitative results analyzed in section 4.1.1, in the top terms, ‘government’, ‘program’,
‘money’, ‘benefits’ and ‘bank’ are the top five most used words in the three month period. In the six month
period, ‘costs’ and ‘children’ displaces ‘bank’ and ‘money’ as the most occurring terms in online
conversations. In the 12 month period, ‘fund’ which could be qualified as ‘money’ surfaces, but would be
disregarded. This means that ‘education’ was one of the most talked about topics.
On the other hand the least talked about topics in descending order, in the three month period are ‘pension
fund’, ‘children’, ‘income’ and ‘republicans’. In the six month period, in descending order, these are; ‘not
get’, ‘student’, ‘interest’, ‘debt’, and ‘pension plan’. While ‘best vacation ever’, ‘federal funding’, ‘plan’,
‘U.S’ and ‘pension plans’ are the least five mentioned topics in the 12 month period.
From the data, the particular area of interest seems to be that of child benefits. This summary is drawn from
the observation that if we look at the three month period and then the six month period, children which was
a least mentioned subject became a top five subject of discussion. This observation also holds for education,
which became a top five topic if a year’s worth of data is considered.
On another front, the five least discussed topics show no associable shifts that could hint at what may be
happening, except for the fact that pension surfaces in the least talked about terms and no topic has moved
from being highly talked about to being a least talked about topic. If this observation of the least talked
about terms holds in the analysis of the top emotions and most used hashtags, only the five top terms will
be the focus of the analysis.
61
Tax: Top terms
62
63
Fig. 18: Graphed Top term representation of the quantified attributes of data from social media networks, associated with Tax
64
During 2012 – 2013, the top term ‘government’ was mentioned more than 1 million times, while the other
four top terms were discussed in about 600,000 conversations during the same period. These numbers are
huge, but a relative comparison to the number of active social media network users on Facebook alone,
brings to mind the reality of modeling errors. This is because about 1 billion people actively use Facebook
each month (The Next Web, Inc., 2013). As such one is right to anticipate that the number of times
‘government’ is mentioned in one year is about 0.12 percent of the active monthly users.
The observation above suggests that, of the 1billion monthly users who use Facebook, that at least the
proportion of comments, tweets, messages or blogs that had conversations relevant to the ‘government’ is
1%. Perhaps until a measure of inaccuracy of the NLP data mining model used in NetBase® can be
quantified, this estimation of 1% contains a degree of uncertainty.
There is some concern regarding the efficacy of the data mining model of NetBase® and whether it is doing
a good job of identifying the right contexts of conversations and extracting most of the relevant topics that
relate to contextually linked keyword another concern is that data on social media networks is evasive of
the data mining model assumptions used, given the pervasive use of diverse slang words in informal
conversations.
Furthermore, on the basis of the postulate that the main subject of discussion is about child benefits, just
like in the qualitative analysis, the data does not tell whether the change made leads to a higher or a lower
support of child benefits from the government. Ordinarily this would in turn give a hint of the likely
direction of its impact on the economic situation of individuals. Therefore, in the absence of such link, it is
appropriate to look at what is happening in the top emotions.
The top five emotions expressed: ‘great’, ‘good’, ‘love’, ‘amazing’ and ‘thank’ all point to a positive
economic disposition about how people felt that year (Fig. 19). On the surface, the ‘top emotions’ seem to
argue the point that whatever changes the government made to tax was a favorable one for families and
consumers at large. Even on a basis of ratios of the most negative emotion expressed relative to the most
positive emotion expressed for the whole year, this gives a ratio of 1:4 and implies that there is a negative
emotion expressed for every four positive emotions expressed.
If the least five positive emotions were taken into account for the whole year (12-month period only), the
words ‘favorite’, ‘proud’, ‘successful’, ‘important’ and ‘vital’ depict an organic picture of a year that has
been highly favorable. Likewise the least five negative expressions for the whole year paint a different
picture. However their frequency is lesser than those of their counterparts in the least five position emotions
expressed. This implies overall, that people’s economic situation was better.
65
Tax: Top emotions expressed
66
67
Fig. 19: Quantified representation of the top emotions expressed that are related to Tax
68
This observation is consistent with the global index of consumer confidence published by Nielsen for the
year 2013. The Nielsen index results estimated there was a 1 percentage point increase in global consumer
confidence (The Nielsen Company, 2014). This does give some credence to the opnion that social media
networks data has a degree of correlation to other established measures of consumer confidence by
renowned organizations like The Nielsen Company.
The other sub category, born in the era of social media networks—the hashtags gives another perspective
to the story. Firstly, in comparison to normal vocabulary expressions, its use is still small. For example for
the period 2012 – 2013, only 25000 of the most used hashtag, ‘#funding’ occurred across social media
networks. It must be noted that hashtags are used predominantly in Twitter, by its 200 million active
monthly users (Jim Edwards - Business Insider, 2013).
In spite of the small scale representation of hashtags relative to the number of active users of Twitter, the
five top hashtags of the year seem to reinforce the postulation that changes to government tax policies were
economically beneficial. And a retrospective comparison between the least top five hashtags and the top
five negative emotions supports drawing such conclusion.
So far the quantitative analysis of the NetBase® data of Table 2, graphed in Fig. 18, Fig. 19 and Fig. 20
illustrates that data from social media networks possess attributes that are good enough for deriving
knowledge of higher order level. This is based on the observation that the data reflects directly what online
conversations are about at each point in time.
Other findings the analysis has uncovered are that hashtags may not be the most effective tool for detecting
what concerns people or is of most interest to them, when it comes to analysis of conversations on social
media networks. The analysis also reflected shifts in online conversations through the movement of
different top terms, as observed in Fig. 18, where it became obvious that tax policies affecting families
changed.
Quantitatively, it is unclear why the volume of the number of terms does not scale with the user activity on
the popular social media networks. But in the face of an assumption that the data mining model implemented
in NetBase® is highly accurate, this reinforces claims that data on social media networks is noisy to an
extent (Daas P.J.H and Puts M.J.H - European Central Bank, 2014).
At this point, further exploration of the attributes of the data from NetBase® is in order. In a similar line of
thought as section 4.1.1, the data for Capital Expenditure will be analyzed in the following subsection.
69
Tax: Most used hashtags
Fig. 20: The hashtags used in conversations that are contextually associated with Tax
70
Capital Expenditure
As an illustration, Fig. 21(c) (p. 72), shows that the five top terms of the year (12 month period), i.e.
‘government’, ‘costs’, ‘market’, ‘tax’ and ‘economy’ are same as the five top terms in the three month and
six month periods, given they all relate to the economic operations of governance. But ‘government’ is by
far the most discussed subject relative to the next top term of the year.
The numbers tell us that the capital expenditure keyword is not as conversational in informal contexts like
the Tax keyword, owing to the frequency of the top term, ‘government’. For instance ‘government’ was
discussed 300,000 times less relative to the number of times it was discussed within the context of the Tax
keyword for the whole of the year (12 month period). The five least top terms on the other hand do not
repeat and so do not give a consistent picture if we look at the entire year’s data.
In comparison to the volume of active users on the social media networks, based on Facebook’s 1billion
active monthly users, the number of times ‘government’ is discussed is a meager 0.08% of this number of
users. Perhaps the fraction of users who actually post comments, messages, update their status would
determine whether this number is indeed small.
Otherwise, this is even more disappointing than that of tax, which happens to be 0.04% higher. And if an
assumption that the data mining model error is insignificant, then this tells us that social media networks
data is indeed noisy. This implies that there is less useful content than intuition makes us believe. Or that
the number of users does not scale with the volume of active usage in form of messages, tweets or blogs.
Since such data is not available, we would assume that posting of comments and online conversations would
be up to 1% at least.
Looking into the ‘top emotions expressed’, Fig. 22(c) for the whole year, both volume and frequency is
very small compared to the number of active users of Facebook’s 1 billion active monthly users. As a ratio
of the total number of active monthly users, this represents 0.00014%. Just like the qualitative analysis
highlighted, there is no hint of how consumers feel. The sub category of the most used hashtags, Fig. 23
does not give this away also.
The analyzes show that the usefulness of data from social media networks rely on the data mining model of
the data collection tool used to extract vocabulary associated with the context of interest. As a result, there
is an assumption that the model embeds any causal relationships between the data relative to the knowledge
domain of interest.
In the next section, the analysis is beamed on the standings of social media networks as sources of data. In
particular the network from which much of the data originated.
71
Capital Expenditure: Top terms
72
73
Fig. 21: The volume of top terms associated with Capital Expenditure
74
Capital Expenditure: Top emotions expressed
75
76
Fig. 22: Graphical representation of the number of Top emotions expressed, relative to Capital Expenditure
77
Capital Expenditure: Most used hashtags
Fig. 23: The most used hashtags related to the subject of Capital Expenditure
78
4.1.3 Part C: An analysis of SMNs content contribution
The analysis done so far has ignored the social media networks, the sources of the data NetBase extracted
for each keyword. The ensuing analysis looks at some these social media networks with the highest number
of conversations that yielded the most relevant data. Then a quantitative analysis compares the number of
conversations to the number of active users on the top two social media networks, to further exploit the
attribute of ‘noisiness’ and data collection ‘model error.’
Figures 25(A) – (G) for the 12 month period, show that Facebook is the social media network with the
highest number of conversations because of the massive size of its network. The keyword Consumer
Confidence is the only exception, where Twitter dominated. This outlier, in some way flags an attribute of
social media networks data that suggests that the use of the data may best be served through a particular
source or group of sources. Across the seven keywords Table 4 below shows the top five social media
networks to buttress this observation.
Facebook.com, twitter.com, reddit.com and scribd.com are the top sources, as seen in Table 4. But their
position in the five top conversation places show that some social media networks could be better sources
with less noise. Of course this depends on the subject of interest. For instance, ‘tax’ presents every traveler
with concerns to do with exchange rate or commodity prices. Businessmen and women whose businesses
are international also face the reality of taxation wherever they do business.
Table 4: Illustration of the how niche social media networks could serve some subject area better than the others
Keywords 1st 2nd 3rd 4th 5th
1. Tax Facebook.com Twitter.com Scribd.com Carefultraveller.com Reddit.com
2. Capital
Expenditure
Faceboook.com Realestateforum.co
m
Twitter.com Scribd.com Reddit.com
3. Health Facebook.com Scribd.com Twitter.com Reddit.com Huffingpost.com
4. Retirement Facebook.com Twitter.com Topix.com Reddit.com Scribd.com
5. Consumer
Confidence
Twitter.com Facebook.com Fontcraft.com Investorshub.advnj.c
om
Online.wsj.com
6. Political
Affiliation
Facebook.com Twitter.com Reddit.com Topix.com Scribd.com
7. Employment
Condition
Facebook.com Twitter.com Scribd.com Blogtalkradio.com Reddit.com
79
realestateforum.com is a strong source of information in the subject of ‘Capital Expenditure; Political
Affiliation—topix.com; Employment condition—blogtalkradio.com and Retirement—topix.com. In fact
the summary of Table 4 is that some social media networks may serve certain purposes better, given their
privileged domain expertise on the subject of interest.
Note, however, that the data in table 4 and in figures 24 a-g does not take into account the number of users
who could potentially be posting on each website, nor does it take into account the total number of messages
posted. We did not have access to such data. If we had, it might have indicated for example that while
carefultraveler.com has many fewer posts than Facebook about Tax, it might still be considered more
influential based on the percentage of users posting or the percentage of posts related to Tax.
Thinking of the fourteen keywords used to extract data from social media networks, it turns out that a social
media network such as Twitter is favorable to the use of phonemes given the 140 character restriction on
the length of a tweet (Twitter Inc., 2014). Whereas an attempt to determine the character limit on Facebook
shows that upwards of 1000 characters is permitted. This implies that a data mining model that is based on
single syllable words would easily process content from Twitter given its briefness than say scribd.com,
which is a social network for electronic books by unpublished authors. This is because the longer the
content, the higher the chances of inaccurately deducing the intentions and actual context of the
conversation.
(a)
Nonetheless, the data shows that Facebook is a dominant influencing source in social media. This is in part
because Facebook permits the posting of story long posts (the length of blogs), short Twitter like posts and
comments by any of its over one billion users. These seem to work in its favor compared to the other social
media networks.
80
Another observation is that the keywords themselves and where most of their data has come from matters.
It is tempting to say that single syllable keywords are best sourced from a social media network like Twitter
and secondary words from other websites that permit a lot of text. This theory could be supported by looking
at Fig. 24(a) for Tax, but is debunked by observing that Facebook is still the major source for other single
syllable keywords like (c) Health and (d) Retirement.
In addition, one could say that keywords that are used frequently in formal conversations about the economy
tend to be announced in the news. As such they are most likely going to get attention on Facebook and
Twitter, the major networks that also supports ‘announcement’ conversations. On the other hand a keyword
that has a perspective of ‘self-interest/learning’ to it is most likely to be found on networks where there are
lengthy, detailed and informed conversations. This is particularly the case with the keyword ‘Health’. As
seen in Fig. 24(c), scribd.com, the online self-publishing service of e-books is the second highest source
after facebook.com (notwithstanding the difference in base audience numbers).
(b)
(c)
81
An analysis of the numbers shows other hidden perspectives about social media network users and what
they feel happy to share. Note that Fig. 24(a) – (g) gives a quantified measure of the number of
conversations that contain each keyword, as opposed to the number of online conversations that took place
on each social media network. On this basis, it could be claimed that topics that are personal and private
are less talked about that those that are external to an individual.
Health (c) and Consumer Confidence (e), seem to be a private affair when compared to Retirement (d) and
Political Affiliation (f). If the top conversation source is considered, Health (c), for example was found in
270,000 conversations on Facebook compared to the 2.3 million on Political Affiliation or 2.8 million on
Tax. More so, Facebook being a very informal network where there are closer interactions between family
and friends, held the most conversations on Health of 270,000, in comparison to the second place network,
scribd.com containing 10,000, even though scribd.com is a prolific [electronic] book open publishing
service with 60 million documents (Scribd., 2015).
(d)
(e)
82
The numbers associated with the number of conversations held on SMNs shed some light on the degree of
how noisy data from SMNs is. In 2014 it was claimed that 510 comments and 293,000 statuses were posted
on Facebook every second (Dan Noyes - Zephora Internet Marketing Solutions, 2014). This results to a
total of ~300,000 conversations per second. Multiplying this number by 60 gives 18 million conversations
per hour. When this is multiplied by 24, it gives an estimation that about 432 million conversations per day.
If night time usage and other cyclic causes of inactivity are taken into account, this number could be say
400 million per day. This means that for a 365-day year, approximately 150 billion conversations were
posted on Facebook.
(f)
(g)
Fig. 24: The top social media networks in terms of the volume of conversations
83
Based on the results of Fig. 24 (a) – (g) which is for the period of 1 year—Oct. 2012 – Oct. 2013, there is
a maximum of 3 million mentions of any one of our keywords on Facebook. This represents 0.002% of the
total number of conversations had on Facebook in one year.
This implies one or more realities as follows: that the conversations on social media networks appear to be
about other subjects that do not in any way relate to the economic situation of users. This hint of
“irrelevance” is questionable because social media networks are used also by governments and businesses
to solicit [subjective] commentary from people. However, caution must be exercised because this does not
imply that the commentary responses are relevant to the topic of CC. Moreover it could be a consequence
of the NLP algorithms not being sufficiently equipped with language rules that cater for the unstructured
language usage on social media network conversations.
Moreover, there is the possibility that the rules and assumptions of the data mining model may not be
aligned with the innate structure of data on social media networks. So in as much as the data may be noisy,
it may also be down to data mining model error. Therefore, it calls to question the need to advance models
that are tenable for deriving higher order levels of knowledge, beyond the literal extraction of phonemes
based on foundational language rules.
At this point in the analysis, the data from social media networks have been explored qualitatively and
examined quantitatively. The sources of the data have also been analyzed and shed insight into some
opportunities for improvement in the methodologies of data mining models. However, there is a case for
examining the characteristics of the raw data that NetBase mined as shown in Table 3(The third output from
a NetBase search; the indexed dataset extracted from social media networks for the search term of Capital
Expenditure.)
4.1.4 Part D: Analysis of raw data indexed by NetBase
The modus operandi of NetBase is that it indexes an enormous amount of data from across social media
networks, thereby making these data available for research at any time. This dictates a reliance on the model
used to mine social media networks for data, which is based on the concept of Natural Language Processing
(NetBase, 2014). Also other factors being relied upon are the frequency of data collection, the metadata
about the data that is indexed and the volume of the data it has to process.
All of the above have a fundamental effect on the data that have been analyzed so far. For instance it affects
the accuracy of the classification of the words associated with the keywords concerning consumer
confidence. The measured frequency of their occurrence, whether it is relevant in context and in meaning
or not, and finally the likelihood of an amplified occurrence due to a high number of repeated raw data
84
being captured more than once due to a short interval between scans of social media networks for data all
influences the usefulness and significance of the data.
Most of the attributes mentioned above can be seen in Table 3 on page 43. The actual posts, tweets or
comments on social media networks are indexed under the column called Sound Bite Text and their
classification as being either negative or positive as Sound Bite Sentiment. If we considered the negative
attribute of repetition it can be seen that rows 10 – 11, 15 – 21 and 13 – 14 of Table 3 are the same. This
represents a total duplication of 7 out of 30 indexed datasets, ~23% noise. For the sake of exploration, let’s
assume that this noise scaled up to the 500 raw datasets, i.e. that 23% of the dataset are repetitions. If we
take the claim of the number of conversations on Facebook per hour as our benchmark, to be 300,000, it
implies 70,000 of those ‘conversations’ are just repetitions not actual conversations on social media
networks. This does throw concern and points to an opportunity for improvement of data mining models.
Furthermore, on the subject of contextual relevance, Table 5 gives a preview into the challenges of
automated classification based on Natural Language Processing. Between the rows of 464 – 485, the raw
sound bite is nothing short of a marketing article espousing the great deal vouchers on Amazon. In no way
do these represent the sentiment of consumers who have used the vouchers and saved some money in the
process. Instead they are just repeated marketing descriptions of vouchers offering good deals. This is the
type of “noise” that was mentioned earlier, and it could be removed to filter the results and improve the
accuracy of the application using the data.
The last attribute of frequency of sampling is examined using Table 6. It is obvious that the sampling
frequency is anything from milliseconds to seconds and sometimes the lower tens of minutes. This is an
incredibly short time owing to the slow pace of time at which human trends evolve, even though human
events happen at a much faster pace. It could be said that this presents opportunities for repetitions but also
for a noisy data that will require filters to get the real data. Nonetheless there is a case for oversampling in
order to have a large enough sample that increases the presence of actual trends.
The case of oversampling is supported by proven concepts from the knowledge domain of Signal processing
in Engineering, under a theory known as The Nyquist-Shannon’s sampling theorem (Weisstein, Eric W. -
From MathWorld -- A Wolfram Web Resource, 2015). It claims that in order to reconstruct a signal, in this
case human expression of sentiment, that all available data on that signal must be sampled at twice the
frequency of the signal of interest. In order words, to measure human events that occur within seconds, we
must sample social media networks at least twice every second to get a picture of the actual events.
Implicitly this would mean that noise will be present in the data as we have already explored.
85
Table 5: Indexed raw data from NetBase associated with Tax for 12 month period
86
Table 6: Part B of Table 5 showing the sampling interval
87
This analysis will be incomplete without a treatment of the data analyzed by another tool to validate the
output received by NetBase in a secondary way. This is covered in the next part, which are the results of
processing the raw data indexed by NetBase in KH Coder, a statistical analysis tool. The aim is to see
whether some fundamental attributes that are assumed to be taken care of by the data mining model used
in NetBase holds true. A particular attribute in this regard is something like the co-occurrence of words.
4.2. KH Coder: Exploration of natural attributes of raw data indexed by NetBase In this section, an examination is undertaken to validate some fundamental rules which ought to hold true
in data analysis. For instance, there ought to be the co-occurrence of words in the English Language
vocabulary or the hierarchical clustering of groups of words based on their associations in a text of the
English Language. Other attributes of the NetBase data are also examined.
The KH Coder application is run on text files constructed from the 500 raw datasets indexed by NetBase.
These represent the period of 12 months. This discussion will focus on just the attributes of the data itself.
Having introduced the co-occurrence analytical method in Chapter 3, it can be seen in Fig. 25 that correct
associations are made between words that normally occur together. On further observation these
associations of co-occurrence center on the Consumer Confidence Index. For example, the Consumer
Confidence Index is usually published in the United States on Tuesday. As the output from KH Coder
shows the words ‘released’ and ‘Tuesday’ (at the bottom of Fig. 25), co-occurred in the dataset contained
the text file KH Coder analyzed.
The word ‘default’, is seen co-occurring with words that are known at the time to be associated with events
of interest as well. For instance, there was a ‘shutdown’ of government services in the U.S due to an
imminent default in honoring the debt obligations of the Federal Government at a time when the
Republicans rejected the proposition made by the Obama administration to raise the borrowing limit. This
co-occurrence shows up in the data
Other words that co-occurred and still do during conversations of economic matters are the British house
price dilemma, the publication of revised consumer confidence index in the U.K.in July and the outlook of
house prices and other sectors of the economy.
In Fig. 26, a non-functional attribute of the raw data set is examined. It shows the relationship between the
frequencies of occurrence of terms (x-axis) with respect to the number of paragraphs (y-axis). The
exponential relationship shown seems obvious, intuitively. Considering that a body of text or population of
88
raw data is about a particular subject of interest, it is inevitable that the volume of data’s growth is a function
of the number of the terms that it contains. That is the interpretation of the result shown in Fig. 26.
Fig. 25: Co-occurrence associations of indexed raw datasets
Fig. 26 highlights the fact that the terms contained in the raw dataset are relevant to the context of the
keyword, albeit the terms sitting at around position 100 have a high density and by implication relevance
to the context of the keyword. By extension they are relevant to the subject of the economic confidence of
consumers.
group
research
plan
forecasting
firm
overall
sample
travel
vacation
Brightens
MJJPEACE
Outlook
U.K.
Christopher
ARTPOP
Flop
Vencace
Will
BoardConference
Chris
Global
IHSInsight
notable
random
quite
deeplysomewhat
perk
release
bitfocus
noticetie
group
research
plan
forecasting
firm
overall
sample
travel
vacation
Brightens
MJJPEACE
Outlook
U.K.
Christopher
ARTPOP
Flop
Vencace
Will
BoardConference
Chris
Global
IHSInsight
notable
random
quite
deeplysomewhat
perk
release
bitfocus
noticetie
group
research
plan
forecasting
firm
overall
sample
travel
vacation
Brightens
MJJPEACE
Outlook
U.K.
Christopher
ARTPOP
Flop
Vencace
Will
BoardConference
Chris
Global
IHSInsight
notable
random
quite
deeplysomewhat
perk
release
bitfocus
noticetie
group
research
plan
forecasting
firm
overall
sample
travel
vacation
Brightens
MJJPEACE
Outlook
U.K.
Christopher
ARTPOP
Flop
Vencace
Will
BoardConference
Chris
Global
IHSInsight
notable
random
quite
deeplysomewhat
perk
release
bitfocus
noticetie
shutdownfunding
speech
speculation
committeedirector
member
FederalReserve
Barnett
House
MoodyWhite
AndrewGoldRock
LTGMr quicker
mainly
realistically
recover*
*
default
negotiatereject
decision
taper
drift
dominateincline
transfix
shutdownfunding
speech
speculation
committeedirector
member
FederalReserve
Barnett
House
MoodyWhite
AndrewGoldRock
LTGMr quicker
mainly
realistically
recover*
*
default
negotiatereject
decision
taper
drift
dominateincline
transfix
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
releaseconfidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
confidence
consumer
price
house
group
research
JulyData
BrightensMJJPEACE
Outlook
U.K.
May
NEW
Reuters
YORK
ARTPOP
Flop
Vencace
WillBoard
Conference
Federal
Reserve
positive
bieberrkfc
british
strongest
rise
add perk
evi
Tuesday
release
89
Fig. 26: Relationship between frequency of terms and the number of paragraphs in a set of data
Another angle from which the raw data set is examined is that of hierarchical clustering. Groups of
associated words that are related to each other relative to the context of the subject matter are structured
into a cluster. Among these clusters is a hierarchy of how they are used in the dataset. Their frequency of
use and number of associations between clusters determines their position on the hierarchy of clusters. This
is shown in Fig. 27.
1 2 5 10 20 50 100 200 500
01
00
20
03
00
40
05
00
Term Frequency
Do
cu
me
nt F
req
ue
ncy (
Pa
rag
rap
hs)
90
Fig. 27: Hierarchical cluster of terms contained in the raw data set indexed by NetBase®
The interpretation relative to the context of consumer confidence does give credence to the fact that terms
such as: confidence, consumer, price, rise, data, house or outlook, feature a lot in the vocabulary used during
formal and informal economic conversations. Whether it is on social media networks or in other forms of
news, this shows the dataset is relevant to the subject of interest—consumer confidence.
confidenceconsumer
positivebieberrkfcbieberrkfc
Positiveprice
riseevi
Julyadd
britishData
houseperkU.K.
OutlookBrightens
MJJPEACEWill
ARTPOPFlop
strengtheyearMay
NEWstrongest
ReutersYORK
Vencacebusinessincrease
datumAugust
economistJunelook
groupresearchTuesdayreleasemarket
Christophersay
BoardConference
gainIndex
expectConsumer
Confidenceweekplan
recentreflect
FederalReserve
newsgood
ratepoint
positivelevel
german%
comegrowth
spendinghit
USnegative
confidenceconsumer
positivebieberrkfcbieberrkfc
Positiveprice
riseevi
Julyadd
britishData
houseperkU.K.
OutlookBrightens
MJJPEACEWill
ARTPOPFlop
strengtheyearMay
NEWstrongest
ReutersYORK
Vencacebusinessincrease
datumAugust
economistJunelook
groupresearchTuesdayreleasemarket
Christophersay
BoardConference
gainIndex
expectConsumer
Confidenceweekplan
recentreflect
FederalReserve
newsgood
ratepoint
positivelevel
german%
comegrowth
spendinghit
USnegative
0 1 2 3 40200400600
91
Chapter 5
Summary & Conclusions
5.1. Summary
The purpose of this thesis has been an exploratory examination of the “sufficiency” of data from social
media networks for the purposes of determining a higher order level of knowledge with a focus on factors
related to the CCI.
A general introduction and history of this index was presented in Chapter 1, highlighting the need for its
improvement with the advent of other sources of data such as social media networks. This approach opens
the feasibility of taking advantage of the fact that obtaining data from social media is timely when compared
to the traditional method that is used to derive the CCI. In this introductory Chapter, the objectives of the
research were set and the research problem was stated—which is to evaluate whether data from social media
networks are “good enough” to measure higher order levels of knowledge.
A review of the body of knowledge and ongoing research effort in the area of computational social science
was reviewed. Chapter 2 presented an in-depth review of published literature in the area of data mining
models and methodologies, where such concepts as NLP and LIWC were explored. It was established that
there was a strong research interest to advance existing models for extracting data from social media
networks for a variety of applications. These focused mainly on micro information synthesis, which falls
short of being applicable for the synthesis of macro information.
In Chapter 3 the methodology for undertaking this research was presented alongside the data collection tool
that was used—NetBase®. In this regard, fourteen keywords frequently used in formal and informal
conversations about the economy and consumer confidence at large were identified. These were in turn
used to search indexed records of conversations from across social media networks stored in NetBase’s
database.
With an outlay of how the research was executed, select keywords were used to gather data from NetBase®
for periods of 3 months, 6 months and 12 months (October 2012 to October 2013). This was done by means
of search, where NetBase® used the input keyword as a reference to extract conversations related to it. Each
search term yielded three outputs from NetBase® (i) a graphical word cloud summary of attributes of a
population of data related to the keyword (ii) a quantified tabular summary of attributes of a population of
92
data related to the keyword and (iii) 500 raw datasets from the data population related to the keyword used
for the search.
These sets of data were presented in Chapter 4 and analyzed to elicit disadvantageous and advantageous
characteristics of data from social media networks. For instance, in order to determine macro knowledge
like the Consumer Confidence Index beyond the exploratory analysis, both qualitative and quantitative, KH
Coder was used to examine some fundamental language rules to examine NetBase’s data collection model
through another lens.
The results from NetBase® were grouped into subcategories, including the Top terms, Top emotions
expressed, most used hashtags, Top attributes, Top behaviors, Top brands and Top dis/likes. ‘Top
behaviors’ was rejected on an analysis category because research showed that humans can better judge
good behavior based on the context than a machine can. The same was for ‘Top attributes’. Other
subcategories were disregarded because they were peripheral to the subject of interest. These include Top
brands, Top likes and Top dislikes.
As a consequence, only the Top emotions expressed, the Top terms and the most used hashtags were the
subcategories examined throughout the analysis. And only seven keywords of the fourteen were analyzed
in order to keep the content of the thesis concise.
The analysis in Chapter 4 shed light on the strengths and weaknesses of data from social media networks.
Chief among its strengths was its reflection of topics of interest that are making the buzz at any one time.
It is also real-time in nature and does not suffer any lags that hinder most forms of traditional data collection,
e.g. CCSs. There is variety and volume when data from social media networks is considered as a source of
information generation, which has been a shortcoming when other forms of data are put into some context
of use beyond their primary domain.
Besides these advantageous attributes of data from social media networks, there are fundamental concerns
around our understanding of the structure of this data. On a fundamental note, there is a knowledge gap in
understanding the structure of this data. This in turn influences the accuracy of the data collection models
which results in a high level of ‘noisiness’ in the data. Such noisiness could be traceable to model error and
also lack of understanding of the nature of social media networks interactions to adjust other factors that
could influence the accuracy of our data collection models. This is important when it has to do with
accurately associating the right context to the intended meaning of a conversation especially with respect
to a keyword of interest.
93
Other findings from the analysis in Chapter 4 also show that the hashtag which is used to make popular
topics of interest easier to ‘trend’ seems irrelevant when it comes down to correctly determining higher
levels of knowledge. Also, the weakness of the hashtag was further exposed, given the popular use of
acronyms and slangs that have no connection with the English Language. In another area, the quantitative
analysis of the results from NetBase® shows that in comparison to the figures being claimed as the number
of monthly active users on a network like Facebook, that the number of conversations relevant to any
particular chosen topic (like ‘tax’) can be relatively small.
Again this reinforces but diversifies the concern of the level of ‘noisiness’ of data from social media
networks and calls into question their sufficiency for macro information synthesis. As was demonstrated,
the highest top term in one of the results occurred in conversations had by only about 0.12 % of Facebook’s
monthly users alone. This is cause for concern, because it implies that either the data mining tool has a
significant ‘model’ error or most conversations on social media networks are grossly irrelevant in the light
of specific named topics of interest.
5.2. Conclusion
The CCI is an influential economic indicator relative to the indicative prospects of an economy. How it is
measured is therefore vital to its continual success as an economic indicator in the near future, especially
as the world’s economy becomes strongly globalized.
Thus far, the dominant traditional method of CCI measurement has proven its reliability, but there is room
for improvement owing to the number of times CCIs are revised after the first publication. An aspect of
CCI measurement that is ripe for improvement is the area of its accuracy, the root cause of the multiple
revisions.
To improve the accuracy of CCIs, it is logical to exploit other sources of data that would complement the
traditional method of evaluating the CCS. This should include only different sources of data, but one whose
data introduces a diversification of perspectives that would give a convergent view on the subject of
interest—CCI. To this end, social media networks was identified as a potential new source of data, given
their proliferation in the past decade and the number of [consumers] who are active users.
Social media networks not only contain data that is rich in variety, but they are a source of data that is
readily accessible in real-time. This data can be extracted within seconds, minutes or hours, as desired. This
is a distinctive advantage over the time it takes to perform a CCS.
94
CCI as a metric quantifies the subjective disposition of consumers about their economic prospects. For this
reason data from social media networks that relate to this topic must be identified correctly, based on the
potential information they give about the economic situation of users holding conversations. This process
of identification and association is critical to the usefulness of any data extracted from social media
networks to develop macro information like a CCI. The tool used to extract this data and the concept of
data mining that such tool implements also affects the usefulness of such data.
The tool used to extract data for this research study from social media networks was NetBase®, a tool which
implements the data mining concept of NLP. Keywords related to the economy, consumers and CCI in
general were used to ‘search’ and ‘extract’ data.
After the data collection was performed, the three types of data collected were analyzed. This analysis
qualitatively showed that data from social media networks has a strong correlation with popular topics of
discussion in the period of time for which data was collected. It also revealed the challenges posed by slangs
and acronyms when the prolific tagging system of hashtag is put under the spotlight. It showed that although
hashtags are associated with making it is easy to identify popular topics that are trending, it does not always
correlate with economic trends.
Furthermore, a quantitative analysis of the data showed that comparatively, the ratio of conversations
deduced to be associated to a keyword (by NetBase®) and the total number of conversations had in the
different social media platforms is not proportional. Consistently this number is small and it gives credence
to the argument that data from social media networks is extremely noisy. It also calls into question whether
data mining concepts such as NLP and their implicit implementation assumption(s) are suitable for the type
of data that trails the unstructured nature of verbal conversations. Perhaps it is suggestive of a need for
further research to focus on adapting these micro information synthesis concepts to serve macro information
synthesis. There is also the need to analyze the increasing use of videos for which there exists little data
mining concepts beyond specific concepts used by engineers to extract intangible information.
The verification analysis performed on the data using KH Coder does confirm that NetBase® and indeed
the NLP concept correctly associate words that are related to a CCI keyword. Words that co-occur do in
fact show a strong correlation in the Co-occurrence graph. The same goes for groups of words that relate
to CCI being identified after a Cluster analytic method was performed in KH Coder. This gives confidence
that at least the NLP concept is adequate for extracting data from social media networks to generate micro
information.
95
In conclusion, this thesis has to some degree revealed that data from social media networks possesses the
attributes of variety and being relevant in real-time. These are surely an advantage over the traditional
method of CCS used to measure CCI. However, when explored further, the issue of correctly associating
the context of conversations with the intentions of the participants arises, relative to the subject of interest
at hand. Therefore, this would induce other forms of error that will make such data unreliable as a single
source of measurement.
Overall, based on the analysis undertaken in this thesis, data from social media networks are “good enough”,
but as only complementary data for use with other forms of data to synthesize higher levels of knowledge.
Therefore, rather than use data from social media networks alone, our advice is that other data formats with
a set scope of context (surveys or interviews), are used as the source of primary data.
5.3. Future Study
This research study exposes numerous areas of work that could improve the future applicability of data
from social media networks to the synthesis of macro information. Starting with the most important aspect
of accuracy of association, concepts that are sufficient for micro information origination need to be
developed further and tested extensively for use cases beyond a micro scope. Additional new concepts that
are better suited to the unstructured form of data from social media networks needs to be developed.
Beyond the level of data association and extraction, methods that could be used to measure whether the
data extracted is related to the keyword (and its context), should be developed. If they are related, there
needs to be a method to measure how many continues to be unrelated and therefore represents the error of
‘model rules’ implemented in a given data mining tool.
Besides these fundamental areas for further research, it would be beneficial to look into emerging language
forms that are unconventional. Examples discovered in this study includes hashtags and hash tagged
acronyms in particular. It is perhaps a right time to seek to understand whether a new form of vocabulary
is emerging, that would enable a better utilization of the insights they provide. It is believed this will
alleviate some of the issues bordering on extraction of irrelevant data from irrelevant conversations.
Finally, it is recommended that a future effort should aim to extract larger data sets from social media
networks, perhaps in the magnitude of thousands to millions. This should be done using a different data
mining model to NLP, with a goal to assess whether the data gathered gives similar hints to the summaries
deduced by NetBase® based on NLP. This would facilitate a comparison between the NLP and such model.
Once this is achieved it will be more possible for an effort to be undertaken to actually derive an accurate
index of Consumer Confidence.
96
Bibliography
Anon., n.d. Office of Research Ethics and Integrity. [Online]
Available at: http://www.research.uottawa.ca/ethics/consent.html
[Accessed 05 April 2015].
Asur, S. & Huberman, B. . A., 2010. Predicting the Future With Social Media. s.l., Web Intelligence and
Intelligent Agent Technology (WI-IAT), 2010 IEEE/WIC/ACM International Conference. IEEE/WIC/ACM,
pp. 492 - 499.
Bank of Ghana, 2011. Bank of Ghana Monetary Policy Report- Real Sector Development. [Online]
Available at:
http://www.bog.gov.gh/privatecontent/MPC_Press_Releases/real%20sector%20indicators%20of%20ec
onomic%20activity%20dec%2011%20-%20reviewedfinal.pdf
[Accessed 17 April 2015].
Barry, Chris; Markey, Rob; Almquist , Eric; Brahm, Chris, 2011. Putting Social Media to Work. [Online]
Available at: http://www.bain.com/Images/BAIN_BRIEF_Putting_social_media_to_work.pdf
[Accessed 17 April 2015].
BBC, 2014. Boom and Bust. [Online] Available at: http://www.bbc.co.uk/bitesize/higher/history/usa/boombust/revision/1/ [Accessed 23 April 2015] Berman, E. & Wang, X., 2011. Essential statistics for public managers and policy analysts. s.l.:CQ Press.
Berry, S. & Davey, M., 2004. How should we think about consumer confidence?. Bank of England
Quarterly Bulletin, Autumn.
Blackshaw, P. & Nazzaro, M., 2004. Consumer-Generated Media (CGM) 101: Word-of-Mouth in the Age
of the Web-Fortified Consumer.. A Neilsen BuzzMetrics White Paper, Spring.
Bloomberg News, April 2003. Consumer Confidence Shows a Substantial Gain. The New York Times, p. 8.
Bogleheads, 2013. Financial Securities. [Online]
Available at: http://www.bogleheads.org/wiki/Financial_securities
[Accessed 20 April 2015]. Casey, G. P. & Owen, A. L., 2013. Good News, Bad News, and Consumer Confidence. Social Science Quarterly, March, 94(1), p. 292–315. Casteleyn, Jonathan, 2013. Major divergence in consumer confidence signals declines for stocks. [Online] Available at: http://marketrealist.com/2013/02/major-divergence-in-consumer-confidence-signals-declines-for-stocks/ [Accessed 20 April 2015].
97
Chaffey, Dave, 2014. Global social media research summary 2014: A compilation of social media
statistics for consumer adoption and usage. [Online]
Available at: http://www.smartinsights.com/social-media-marketing/social-media-strategy/new-global-
social-media-research/
[Accessed 20 April 2015].
Charles Arthur, The Guardian, 2014. Facebook emotion study breached ethical guidelines, researchers
say. [Online]
Available at: http://www.theguardian.com/technology/2014/jun/30/facebook-emotion-study-
breached-ethical-guidelines-researchers-say
[Accessed 21 April 2015].
Chitwan, R., 2002. The Life History Calendar as a Data Collection Tool. [Online] Available at: http://perl.psc.isr.umich.edu/Res_Brief_2_final.pdf [Accessed April 2015]. Collobert, R. et al., 2011. Natural Language Processing (Almost) from Scratch. Journal of Machine Learning Research, Volume 12, pp. 2461-2505. Cotsomitis, J. A. & Kwan, A. C. C., 2006. Can Consumer Confidence Forecast Household Spending? Evidence from the European Commission Business and Consumer Surveys. Southern Economic Journal, January, 72(3), pp. 597-610. Curtin, R. T., 2002. Surveys of Consumers: Theory, Methods, and Interpretation. Washington DC, s.n.
Daas, P. J. & Puts, M. J., 2014. Social Media Sentiment and Consumer Confidence. [Online]
Available at: https://www.ecb.europa.eu/pub/pdf/scpsps/ecbsp5.pdf
[Accessed April 2015].
Dan Noyes - Zephora Internet Marketing Solutions, 2014. The Top 20 Valuable Facebook Statistics -
Updated October 2014. [Online]
Available at: https://zephoria.com/social-media/top-15-valuable-facebook-statistics/
[Accessed 12 April 2015].
Dominitz, J. & Manski, C. F., 2004. How Should We Measure Consumer Confidence?". Journal of
Economic Perspectives, Spring, 18(2), pp. 51-66.
Elizabeth Dwoskin, The wall Street Journal , 2013. Twitter's Data Business Proves Lucrative. [Online]
Available at: http://online.wsj.com/news/articles/SB10001424052702304441404579118531954483974
[Accessed 12 April 2015].
Elson, S. B. et al., 2012. Using Social Media to guage Iranian Public Opininon and Mood After the 2009
Election. [Online]
Available at: http://www.rand.org/content/dam/rand/pubs/technical_reports/2012/RAND_TR1161.pdf
[Accessed April 2015].
98
Encyclopædia Britannica, 2015. Social status. [Online]
Available at: http://www.britannica.com/EBchecked/topic/551450/social-status
[Accessed 07 April 2015].
Erik Qualman, 2009. Statistics Show Social Media Is Bigger Than You Think. [Online]
Available at: http://www.socialnomics.net/2009/08/11/statistics-show-social-media-is-bigger-than-you-
think/
[Accessed 03 April 2015]
Fredrickson, B. L. & Losada, M. F., 2005. Positive Affect and the Complex Dynamics of Human
Flourishing. American Psychologist, October, 60(7), pp. 678-686.
Freedman, D. et al., 1988. The Life History Calendar: A Technique for Collecting Retrospective Data.
American Sociological Association, Volume 18, pp. 37-68.
FxPro Blog, 2014. U.S. Michigan consumer sentiment index revised down to 88.8. [Online]
Available at: http://blog.fxpro.co.uk/u-s-michigan-consumer-sentiment-index-revised-down-to-88-8/
[Accessed 20 April 2015].
Garner, A. C., 1981. Economic determinants of consumer sentiment. Journal of Business Research, 9(2),
p. 205–220.
Gillin, P. & Schwartzman, E., 2010. Social marketing to the business customer: Listen to your B2B
Market, generate major account leads, and build client relationships. In: Social marketing to the business
customer. illustrated ed. s.l.:John Wiley & Sons, p. 256.
GNIP, 2010. How Does Netbase Achieve the Best Accuracy for Understanding Consumers Online?.
[Online]
Available at: http://gnip.com/docs/Listening_Accuracy_White_Paper.pdf
[Accessed 14 April 2015].
Gritten, A., 2011. New insights into consumer confidence in financial services. International Journal of
Bank Marketing, 29(2), pp. 90 - 106.
Higuchi, K., 2013. Cluster Analysis: Indicate Clusters by Different Colors. [Online]
Available at: http://khc.sourceforge.net/en/gallery/pages/image/imagepage6.html
[Accessed 07 April 2015].
Higuchi, K., 2013. Co-occurrence Network: Betweenness Centrality. [Online]
Available at: http://khc.sourceforge.net/en/gallery/pages/image/imagepage10.html
[Accessed 18 April 2015].
Higuchi, K., 2013. Correspondence Analysis. [Online]
Available at: http://khc.sourceforge.net/en/gallery/pages/image/imagepage15.html
[Accessed 07 April 2015].
99
Higuchi, K., 2014. KH Coder. [Online] Available at: http://khc.sourceforge.net/en/ [Accessed 07 April 2015] Hogenboom, A. et al., 2010. Mining Economic Sentiment using Argumentation Structures. s.l., In Advances in Conceptual Modeling- Applications and Challenges. Springer Berlin Heidelberg, pp. 200-209. Hopkins, D. J. & King, G., 2010. A method of automated nonparametric content analysis for social science. American Journal of Political Science, 54(1), pp. 229-247. Hudszinki, D., 2014. A private discussion on Nielsen's market intelligence product for retail: Consumer Confidence monitoring [Interview] [12 December 2014]. Hulbert, Mark, 2010. It's darkest before the dawn- Commentary: Bull markets often begin before consumer confidence hits bottom. [Online] Available at: http://www.marketwatch.com/story/contrarian-view-on-consumer-confidence-2010-02-24?reflink=MW_news_stmp [Accessed 21 April 2015]. Hymans, S. H., Ackley , G. & Juster, T. F., 1970. Consumer Durable Spending: Explanation and Prediction. Brookings Papers on Economic Activity, Volume 2, pp. 173-206. IBM, 2014. What is Watson? As a cognitive technology, Watson is a natural extension of what humans can do at their best. [Online] Available at: http://www.ibm.com/smarterplanet/us/en/ibmwatson/what-is-watson.html [Accessed 20 April 2015]. INSEAD, 2009. How Obama used social networking tools to win. [Online] Available at: http://knowledge.insead.edu/innovation/how-obama-used-social-networking-tools-to-win-1600 [Accessed 17 April 2015]. INSEE- National Institute of Statistics and Economic Studies, 2004. Monthly consumer confidence survey. [Online] Available at: http://www.insee.fr/en/themes/indicateur.asp?id=20 [Accessed 02 April 2015]. Investopedia, 2013. Present Situation Index. [Online] Available at: http://www.investopedia.com/terms/p/presentsituationindex.asp [Accessed 14 April 2015]. Investopedia LLC, 2015. Complete Guide to Corporate Finance. [Online] Available at: http://www.investopedia.com/walkthrough/corporate-finance/2/taxes/types-taxes.aspx [Accessed 18 April 2015]. Investopedia, LLC, 2015. Capital Expenditure - CAPEX. [Online] Available at: http://www.investopedia.com/terms/c/capitalexpenditure.asp [Accessed 14 April 2015].
100
Investopedia, 2013. Expectations Index. [Online] Available at: http://www.investopedia.com/terms/e/expectationsindex.asp [Accessed 14 April 2015]. Investopedia, 2014. Capital Investment. [Online] Available at: http://www.investopedia.com/terms/c/capital-investment.asp [Accessed 14 April 2015]. Jansen, W. & Nahuis, N. J., 2003. The stock market and consumer confidence: European evidence. Elsevier, April, Issue 1, pp. 89-98. Jeremy M. Piger, 2003. Consumer Confidence Surveys: Do They Boost Forecasters' Confidence?. [Online] Available at: https://www.stlouisfed.org/publications/re/articles/?id=415 [Accessed 11 April 2015]. Jim Edwards - Business Insider, 2013. Twitter's 'Dark Pool': IPO Doesn't Mention 651 Million Users Who Abandoned Twitter. [Online] Available at: http://www.businessinsider.com/twitter-total-registered-users-v-monthly-active-users-2013-11?IR=T [Accessed 19 April 2015]. Jurafsky, D. & Manning, C., 2014. Natural Language Processing. [Online] Available at: https://class.coursera.org/nlp/lecture/124 [Accessed 06 April 2015]. Karen Christensen - Forbes India, 2014. Incidental Vs. Integral: Understanding your Emotions. [Online] Available at: http://forbesindia.com/article/rotman/incidental-vs.-integral-understanding-your-emotions/36949/1 [Accessed 11 April 2015]. Katona, G., 1960. The powerful consumer, s.l.: McGraw-Hill. Keller Fay Group, 2010. Four WOM Statistics. [Online] Available at: http://www.kellerfay.com/four-wom-statistics/ [Accessed 3 April 2015]. Kongthon , A., Sangkeettrakarn , C., Kongyoung , S. & Haruechaiyasak , C., 2009. Implementing an online help desk system based on conversational agent. New York, Proceedings of the International Conference on Management of Emergent Digital EcoSystems. ACM, p. 69. Kramer, D., Guillory, J. & Hancock, J., 2014. Experimental evidence of massive-scale emotional contagion
through social networks. Proceedings of the National Academy of Sciences of the United States of
America, 25 03, pp. 8788-8790.
Kwan, A. C. C. & Cotsomitis, J. A., 2006. The Usefulness of Consumer Confidence in Forecasting
Household Spending in Canada: A National and Regional Analysis. Economic Inquiry, 44(1), pp. 185-197.
Lemmens, A., Croux, C. & Dekimpe, M. G., 2007. Consumer Confidence in Europe: United in diversity?.
International Journal of Research in Marketing, June, 24(2), pp. 113-127.
101
Liu, B., 2010. Sentiment Analysis: A Multi-Faceted Problem. IEEE Intelligent Systems, 25(3), pp. 76-80.
Lovell, M. C., 1975. Why was the Consumer Feeling So Bad?. Brookings Papers on Economic Activity,
Volume 2, pp. 473-479.
Ludvigson, S. C., 2004. Cosumer Confidence and Consumer Spending. Journal of Economic Perspectives,
Spring, Volume 18, pp. 29-50.
Madigan, Kathleen ;, 2015. U.S. Consumer Confidence Surges to Highest Level Since 2007. [Online]
Available at: http://www.wsj.com/articles/u-s-consumer-confidence-surges-to-highest-level-since-2007-
1422371844
[Accessed 17 April 2015].
Mangold, G. W. & Faulds, D. J., 2009. Social media: The new hybrid element of the promotion mix.
Elsevier, 52(4), p. 357–365.
MasterCard Worldwide, 2010. MasterCard Worldwide Index of Consumer Confidence. [Online]
Available at:
http://www.masterintelligence.com/view_report/index_of_consumer_confidence/africa/h1_2010/Nige
ria
[Accessed 14 April 2015].
McKinsey & Company, 2012. The social economy: Unlocking value and productivity through social
technologies. [Online]
Available at: http://www.mckinsey.com/insights/high_tech_telecoms_internet/the_social_economy
[Accessed 11 April 2015].
Merkle, D. M., Langer, G. E. & Sussman, D., 2004. Consumer Confidence: Measurement and Meaning.
New York, NY, ABC News.
Microsoft Research, 2014. Natural Language Processing. [Online]
Available at: http://research.microsoft.com/en-us/groups/nlp/
[Accessed 12 April 2015].
Mishne, G. & Glance, N., 2006. Predicting Movie Sales from Blogger Sentiments. American Association
for Artificial Intelligence, Spring Symposium: Computational Approaches to Analyzing Weblogs, pp. 155-
158.
Mowen, J. C., 1999. The 3M Model of Motivation and Personality-Theory and Emperical Applications to
Consumer Behaviour. Massachusetts: Kluwer Academic Publishers.
Mueller, E., 1963. Ten Years of Consumer Attitude Surveys: Their Forecasting Record. Journal of the
American Statistical Association, December, Volume 58, pp. 899-917.
Mueller, E., 1963. Ten Years of Consumer Attitude Surveys: Their Forecasting Record. Journal of the
American Statistical Association, 58(304), pp. 899-917.
102
Naoi, T., Yamada, T., Iijima, S. & Kumazawa, T., 2011. Applying the caption evaluation method to studies
of visitors’ evaluation of historical districts. Tourism Management, Volume 32, pp. 1061-1074.
National Restaurant Association, 2014. State of the American Consumer. [Online]
Available at: http://www.restaurant.org/News-Research/News/State-of-the-American-consumer
[Accessed 18 April 2015].
Netbase Solutions, 2012. Social Intelligence Warehouse Drives Instant Answers and Action. [Online]
Available at: http://www.netbase.com/social-intelligence/consumer-base/
[Accessed 02 April 2015].
Netbase, 2010. How Does Netbase Achieve the Best Accuracy for Understanding Consumers Online?.
Netbase, September.
Available at: https://gnip.com/docs/Listening_Accuracy_White_Paper.pdf Accessed 23 April 2015 Netbase, 2012. [Online] Available at: http://www.netbase.com/ [Accessed 01 April 2015]. NetBase, 2014. Why Use NetBase?. [Online]
Available at: http://www.netbase.com/social-media-management/why-use-netbase/
[Accessed 01 April 2015].
Nielsen, 2013. Consumer COnfidence Concerns and Spending Intentions around the World. [Online]
Available at: http://www.nielsen.com/us/en/insights/reports/2014/consumer-confidence-concerns-
and-spending-intentions-around-the-world.html
[Accessed 11 April 2015]. Nielsen, 2013. Global Consumer Confidence. [Online] Available at: http://www.nielsen.com/us/en/nielsen-solutions/nielsen-measurement/global-consumer-confidence.html [Accessed 11 April 2015]. Nielson, 2013. Consumer COnfidence Concerns and Spending Intentions around the World. [Online] Available at: http://www.nielsen.com/us/en/insights/reports/2014/consumer-confidence-concerns-and-spending-intentions-around-the-world.html [Accessed 11 April 2015]. Nielson, 2014. Global Consumer Confidence. [Online] Available at: http://www.nielsen.com/us/en/solutions/measurement/global-consumer-confidence.html [Accessed 11 April 2015]. Nigam, K. & Hurst, M., 2004. Towards a Robust Metric of Opinion. In AAAI Spring Symposium on Exploring Attitude & Affect in Text. [Online] Available at: http://www.kamalnigam.com/papers/metric-EAAT04.pdf [Accessed 08 April 2015].
103
Nyu.edu, 2014. Consumer Confidence. [Online] Available at: http://pages.stern.nyu.edu/~nroubini/bci/ConsumerConfidence.html [Accessed 08 April 2015]. O'Connor, B., Balasubramanyan, R., Routledge, B. R. & Smith, N. A., 2010. From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series. ICWSM, Volume 11, pp. 122-129. Oest, R. v. & Franses, H. P., 2008. Measuring changes in consumer confidence. Journal of Economic Psychology, June, 29(3), pp. 255-275. Office of National Statistics, 2013. Social Networking: The UK as a Leader in Europe, United Kingdom: Office of National Statistics. [Online] Available at: http://www.ons.gov.uk/ons/rel/rdit2/internet-access---households-and-individuals/social-networking--the-uk-as-a-leader-in-europe/sty-social-networking-2012.html [Accessed 18 April 2015] Olowofeso, O. E. & Doguwa, S., 2012. Consumer Sentiment and Confidence Indices in Nigeria: A Panel Data Analysis. [Online] Available at: http://www.bis.org/ifc/events/6ifcconf/olowofesodoguwa.pdf [Accessed 18 April 2015]. Oppewal, H., Paas, L. J., Crouch, G. I. & Huybers, T., 2010. Segmenting consumer based on how they spend a tax rebate: An analysis of the Australian stimulus payment. Journal of Economic Psychology, August, 31(4), pp. 510-519. Otoo, M. W., 1999. Consumer Confidence and the Stock Market. Board of Governors of the Federal
Reserve System Research Paper Series - FEDS Papers, November.pp. 99-60 .
Out:Think, n.d. The 6 types of Social Media. [Online]
Available at: http://outthinkgroup.com/tips/the-6-types-of-social-media
[Accessed 01 April 2015].
Oxford Dictionaries, 2015. Oxford Dictionaries - Language matters. [Online]
Available at: http://www.oxforddictionaries.com/
[Accessed 06 April 2015].
Pak, A. & Paroubek, P., 2010. Twitter as a Corpus for Sentiment Analysis and Opinion Mining. LREC, May,
Volume 10, pp. 1320-1326.
Pennebaker, J. W., Booth, R. J. & Francis, M. E., 2007. Linguistic Inquiry and Word Count: LIWC2007.
[Online]
Available at: http://psycholingua.googlecode.com/svn-
history/r79/trunk/resources/LIWC2007_OperatorManual.pdf
[Accessed 20 April 2015].
Performics, 2010. Social Networking Study: Facebook Use Continues to Rise; Brand Participation and
Engagement Heavily Welcomed by Social Networkers. [Online]
104
Available at: http://www.performics.com/social-networking-study-facebook-use-continues-to-rise-
brand-participation-and-engagement-heavily-welcomed-by-social-networkers/
[Accessed 12 April 2015].
PR Newswire, 2013. The Conference Board's Consumer Confidence Index rebounds in April, s.l.: PR
Newswire Association LLC.
Research at Google, Google Inc., 2014. Natural Language Processing. [Online]
Available at: http://research.google.com/pubs/NaturalLanguageProcessing.html
[Accessed 19 April 2015].
Reuters, 2014. U.S. consumer confidence falls in November to lowest since June. [Online]
Available at: http://www.reuters.com/article/2014/11/25/us-usa-economy-confidence-
idUSKCN0J91O920141125
[Accessed 06 April 2015].
Roberts, I. & Simon, J., 2001. What Do Sentiment Surveys Measure?. [Online]
Available at: http://www.rba.gov.au/publications/rdp/2001/pdf/rdp2001-09.pdf
[Accessed 20 April 2015].
Rouse, Margaret, 2011. Capex (capital expenditure). [Online] Available at: http://whatis.techtarget.com/definition/CAPEX-capital-expenditure [Accessed 11 April 2015]. Rouse, Margaret, 2014. Social Media. [Online] Available at: http://whatis.techtarget.com/definition/social-media [Accessed 11 April 2015]. Ryan Barnes, Investopedia, 2009. Economic Indicators: Consumer Confidence Index (CCI). [Online] Available at: http://www.investopedia.com/university/releases/consumerconfidence.asp [Accessed 02 April 2015]. Ryan, J. - Bloomberg, 2014. U.K. Consumer Confidence Stalls as Economic Outlook Deteriorates. [Online] Available at: http://www.bloomberg.com/news/2014-11-28/u-k-consumer-confidence-stalls-as-economic-outlook-deteriorates.html [Accessed 06 April 2015]. Schweidel, D. A., Moe, W. W. & Boudreaux, C., 2012. Social Media Intelligence: Measuring Brand Sentiment from Online Conversations. Marketing Science Institute, pp. 12-100. Scribd., 2015. About - The word's favourite open publishing platform. [Online]
Available at: https://www.scribd.com/about
[Accessed 11 April 2015].
Sergie, M.A.; Kayakiran, F.; Bloomberg, 2015. Qatar, Shell Scrap $6.5 Billion Project After Oil's Drop.
[Online]
Available at: http://www.bloomberg.com/news/2015-01-14/qatar-shell-scrap-6-5-billion-project-amid-
105
oil-price-collapse.html
[Accessed 16 April 2015].
Sr. Bogue, Edith;, 2008. Correlation and Regression SPSS. [Online]
Available at: http://www.slideshare.net/edithosb/correlation-and-regression-spss
[Accessed 20 April 2015]. Statista, 2014. Statistics and Market Data on Social Media & User-Generated Content. [Online] Available at: http://www.statista.com/markets/424/topic/540/social-media-user-generated-content/ [Accessed 20 April 2015]. Tausczik, Y. R. & Pennebaker, J. W., 2010. The Psychological Meaning of Words: LIWC and Computerized
Text Analysis Methods. Journal of Language and Social Psychology, 29(1), p. 24–54.
The Conference Board of Canada, 2014. Consumer Confidence. [Online]
Available at: http://www.conferenceboard.ca/topics/economics/consumer_confidence.aspx
[Accessed 06 April 2015].
The Conference Board of Canada, 2014. Consumer Confidence US. [Online]
Available at: http://www.conferenceboard.ca/topics/economics/consumer-confidence-us.aspx
[Accessed 06 April 2015].
The Conference Board, 2011. Consumer Confidence Survey Technical Note - February 2011. [Online]
Available at: https://www.conference-board.org/pdf_free/press/TechnicalPDF_4134_1298367128.pdf
[Accessed 06 April 2015] The Conference Board, 2011. Consumer Confidence Survey® Technical Note – February 2011. [Online] Available at: http://www.conference-board.org/pdf_free/press/TechnicalPDF_4134_1298367128.pdf [Accessed 06 April 2015]. The Conference Board, 2014. The Conference Board Consumer Confidence Index® Declines. [Online] Available at: http://www.conference-board.org/press/pressdetail.cfm?pressid=5397 [Accessed 12 April 2015]. The New York State Department of Taxation and Finance, 2014. More Tax Types. [Online]
Available at: http://www.tax.ny.gov/forms/more_tax_types.htm
[Accessed 28 April 2015].
The Next Web, Inc., 2013. Facebook passes 1.23 billion monthly active users, 945 million mobile users,
and 757 million daily users. [Online]
Available at: http://thenextweb.com/facebook/2014/01/29/facebook-passes-1-23-billion-monthly-
active-users-945-million-mobile-users-757-million-daily-users/
[Accessed 21 April 2015].
The Nielsen Company, 2014. Nielson: Global Consumer Confidence Holds Steady to Close 2013; While
Spending Intentions Retreat to Q4 2012 Levels. [Online]
Available at: http://www.nielsen.com/us/en/press-room/2014/nielsen-global-consumer-confidence-
106
holds-steady-to-close-2013.html
[Accessed 04 April 2015].
The Wall Street Journal, 2013. The Oracle of Cyberspace?- The Online Buzz about Warren Buffet joining
Twitter. [Online]
Available at:
http://online.wsj.com/article/SB10001424127887324766604578459392169116764.html?mod=WSJ_Bo
oks_LS_Books_5
[Accessed 25 April 2015].
The Wall Street Journal, 2014. November U.S. Consumer Confidence Unexpectedly Falls to 88.7. [Online]
Available at: http://online.wsj.com/articles/november-u-s-consumer-confidence-unexpectedly-falls-to-
88-7-1416928693
[Accessed 25 April 2015].
Thelwall, M., Buckley, K. & Paltoglou, G., 2011. Sentiment in Twitter Events. Journal of the American
Society for Information Science and Technology, 62(2), p. 406–418.
Thomson Reuters & University of Michigan, 2014. Surveys of consumers. [Online]
Available at: http://data.sca.isr.umich.edu/fetchdoc.php?docid=50688
[Accessed 14 April 2015].
Tobin, L., 2012. Entrepreneur: How to Start an Online Business. 1st ed. Chichester: Wiley & Sons.
TradingEconomics, 2014. United Kingdom Consumer Confidence. [Online]
Available at: http://www.tradingeconomics.com/united-kingdom/consumer-confidence
[Accessed 28 April 2015]
Tsytsarau, M. & Palpanas, . T., 2012. Survey on mining subjective data on the web. Data Mining and
Knowledge Discovery, 24(3), pp. 478-514.
Tumasjan, A., Sprenger, T. O., Sandner, P. G. & Welpe, I. M., 2010. Predicting Elections with Twitter:
What 140 Characters Reveal about Political Sentiment. s.l., ICWSM, pp. 178-185.
Twitter Inc., 2014. Character Counting. [Online]
Available at: https://dev.twitter.com/overview/api/counting-characters
[Accessed 21 April 2015].
Twitter, Inc., 2014. Using hashtags on Twitter. [Online]
Available at: https://support.twitter.com/articles/49309-using-hashtags-on-twitter#
[Accessed 21 April 2015].
unesco.org, 2008. Correspondence Analysis. [Online]
Available at: http://www.unesco.org/webworld/idams/advguide/Chapt6_5.htm
[Accessed 7 April 2015].
107
United Nations, 2015. World Economic Situation and Prospects 2015. [Online]
Available at: http://www.un.org/en/development/desa/publications/wesp-2015-2.html
[Accessed 3 April 2015].
Vanguardngr, 2011. MasterCard ranks Nigeria most optimistic market in Africa. [Online]
Available at: http://www.vanguardngr.com/2011/08/mastercard-ranks-nigeria-most-optimistic-market-
in-africa/
[Accessed 3 April 2015]
Weisstein, Eric W. - From MathWorld -- A Wolfram Web Resource, 2015. Sampling Theorem. [Online]
Available at: http://mathworld.wolfram.com/SamplingTheorem.html
[Accessed 12 April 2015].
